Skip to content

For years, recovery planning followed a familiar pattern. Build the plan, document the steps, and assume it will work when needed. For a long time, that approach held up. Hardware failures, isolated outages, even natural disasters – these were scenarios organizations could anticipate and plan for with some level of confidence.

But the equation has changed.

In this episode of STRIVE, I sat down with Commvault’s Jason Cray, Principal Product Experience, to explore a reality we continue to see across organizations of all sizes: Most don’t fail because they lack a recovery plan. They fail because they’ve never proven that plan will hold up under real pressure.

Watch the full episode.

Key Takeaways: Why Recovery Plans Break Down

  • A documented plan isn’t the same as a proven one. If it hasn’t been tested in realistic conditions, it’s still an assumption.
  • Recovery is a team sport. Security, infrastructure, and operations must align – or recovery slows down.
  • Most investment still happens “left of boom.” Prevention matters, but recovery readiness often gets overlooked.
  • Testing exposes gaps and builds confidence. Without it, organizations default to hope.
  • Resilience is an operational discipline. It requires iteration, communication, and continuous improvement.

The Problem With ‘It Should Work’

On paper, recovery looks straightforward. You define when to recover to, what needs to come back, and where it should be restored. The process appears logical, structured, and manageable.

But as Jason points out, that simplicity rarely survives real-world conditions.

Plans are written in controlled environments, but they’re executed in chaos. When an incident hits, teams aren’t calmly stepping through documentation – they’re reacting, troubleshooting, and trying to align in real time. That’s where the gap emerges. Not between tools and technology, but between expectation and execution.

Sneak Peek: Why Plans Fail Under Pressure

In this moment from the conversation, Jason and I break down why having a plan isn’t enough – and what it actually takes to know a plan will work when it matters.

We’ve Seen This Before

What’s interesting is that this isn’t a new problem; it’s a familiar one, just in a different context.

If you go back to the early days of disaster recovery, organizations followed a similar pattern. Plans existed, but testing was inconsistent at best. Jason shared an example of spending an entire night helping a client pass a disaster recovery test they thought they were ready for. The plan looked solid. The execution told a different story.

Over time, organizations adapted. They tested more frequently, introduced failover exercises and, in some cases even ran production from secondary environments to prove readiness. That shift from assumption to validation is exactly what cyber resilience now requires.

The First Breakdown: Communication

If there’s one issue that consistently surfaces, it’s communication.

In many organizations, responsibilities are clearly defined – security handles prevention, infrastructure manages systems, and operations owns recovery. Individually, each team may be doing exactly what they’re supposed to do.

But recovery doesn’t happen in isolation. It depends on how well those teams work together when something goes wrong.

As Jason describes, too often it becomes a handoff model: “We’ve done our part, now it’s someone else’s turn.” That approach introduces delays, confusion, and ultimately risk. During a cyber event, coordination matters more than ownership.

The ‘Left of Boom’ Problem

Another pattern we continue to see is the imbalance in where organizations focus their efforts.

There’s significant investment in prevention – security tools, detection platforms, and defensive strategies designed to stop an attack before it happens. That investment is necessary, and it plays a critical role.

But far less attention is given to what happens after the event.

The assumption is that if enough effort is spent on prevention, recovery becomes a secondary concern. In reality, the opposite is true. At some point, something gets through. And when it does, recovery becomes the defining factor in how an organization responds.

From Hope to Evidence

This is where the mindset needs to shift.

It’s not about adding more tools or rewriting documentation. It’s about moving from a model based on hope to one grounded in evidence.

Jason highlights a key observation: The organizations that handle disruption well aren’t the ones that avoid incidents – they’re the ones that experience less impact when those incidents occur. They’ve tested their processes. They’ve validated their assumptions. They understand where their gaps are.

Most importantly, they’ve built confidence – not by believing the plan will work, but by proving it.

Start Small, Build Momentum

For many teams, the challenge isn’t understanding the problem – it’s knowing where to begin.

The answer isn’t to overhaul everything at once. It’s to start small and build from there.

Focus on one or two critical services. Understand what’s required to recover them. Bring together the teams responsible for those systems and test the process end-to-end. From there, expand the scope and continue refining.

This approach does more than improve recovery – it builds alignment, reinforces communication, and creates the foundation for broader resilience.

The Reality: No Plan Survives First Contact

One of the most honest moments in our discussion was this: Even the best plan won’t work exactly as written.

That’s not a failure – it’s expected.

Jason puts it simply: If you don’t have a plan, you will fail. But even if you do have one, it won’t unfold perfectly in the moment.

What matters is how prepared to adapt your teams are. Testing creates that adaptability. It builds the muscle memory needed to respond effectively when conditions don’t match expectations.

Watch the Full Episode

There’s much more we cover in this STRIVE conversation, including:

  • Why recovery plans often fail despite being well-documented.
  • What differentiates organizations that recover effectively.
  • How communication gaps impact execution.
  • Where to start when improving recovery readiness.
  • Why testing is the foundation of resilience.

Watch now.

If you’ve ever questioned whether your recovery plan would actually work, this is a conversation worth your time.

FAQs

Q: Why isn’t having a recovery plan enough?

A: Because most plans are never validated under real-world conditions. Without testing, they remain assumptions rather than proven strategies.

Q: What causes recovery plans to fail?

A: The most common issues in recovery plans are lack of testing, poor cross-team communication, and gaps between documented processes and real execution.

Q: What does “left of boom” mean?

A: Left of boom refers to the focus on preventing incidents before they occur. Many organizations invest heavily here but underinvest in recovery capabilities.

Q: How often should recovery plans be tested?

A: Recovery plans should be tested regularly and under varied conditions. Testing should simulate realistic scenarios, not just controlled exercises.

Q: Where should organizations start?

A: Start with a small set of critical services, align the responsible teams, and test recovery end-to-end before expanding.

Q: What is the key mindset shift?

A: Moving from hope-based planning to evidence-based validation.

Chris Mierzwa is Senior Director, Portfolio Marketing, at Commvault.

More related posts


Thumbnail_Blog-Testing-Once-a-Year-2026

Testing Once a Year Is Not a Resilience Strategy

Read more about Testing Once a Year Is Not a Resilience Strategy
Thumbnail_Blog-IDC-Resops-2026

From Recovery to ResOps™: Building Enterprise Resilience That Scales

Read more about From Recovery to ResOps™: Building Enterprise Resilience That Scales
Readiverse-Featured-Image-888-x-500

Ready Is Good. Resilient Is Better.

Read more about Ready Is Good. Resilient Is Better.

We’ve spent years focusing on identity security in the context of people – who have access, what they can do, and how to control it. That model made sense when most activity in the environment was driven by human users.

But that’s no longer the case.

Machine identities – applications, services, APIs, and automated workloads – now play a central role in how modern systems operate. They authenticate, communicate, and execute tasks, often without direct oversight. And in many environments, they already outnumber human identities by a wide margin.

In this episode of STRIVE, I sit down with Dan Conrad, Principal Technologist and a fellow Field CTO at Commvault. We take a closer look at what that shift means – not just from a security perspective, but also from a governance standpoint. And we explore why so many organizations are still treating this as a secondary concern.

Watch the full episode.

Key Takeaways: Where the Risk Is Shifting

  • Machine identities are growing faster than human identities, often by orders of magnitude.
  • Governance models haven’t kept pace, creating blind spots in access and control.
  • Visibility is the core challenge. Many teams don’t fully understand how machine identities behave.
  • Privilege sprawl extends beyond users, with machine identities often holding persistent access.
  • Resilience depends on understanding and managing the remit of these machine identities before they become a problem.

The Identity Model Has Changed

For a long time, identity management was relatively straightforward. You could map users to roles, define access policies, and build controls around predictable behavior. Even with complexity, the model was still anchored in human activity.

Machine identities have broken that model.

They’re created dynamically, often as part of development or deployment processes. They interact across systems in ways that aren’t always visible or well-documented or audited. And unlike human users, they don’t follow a clean lifecycle – they aren’t onboarded and offboarded in the same structured way.

That creates a different kind of challenge. It’s not just about controlling access anymore. It’s about understanding how that access is being used, how it evolves, and how it connects across the environment.

Sneak Peek: You Can’t Phish a Non-Human Identity

In this moment from the STRIVE discussion, Dan describes how attackers aren’t targeting non-human identities directly through phishing – they’re malicious actors using compromised human accounts through social engineering, as a steppingstone to escalate privileges and impersonate powerful machine identities. Once inside, techniques like pass-the-hash and overprivileged service accounts allow attackers to move laterally and vertically, even after passwords are reset.

The Governance Gap

The real issue isn’t that machine identities exist , it’s how they’re governed.

In most organizations, there’s a clear process for managing human access:

  • Requests are approved.
  • Permissions are reviewed.
  • Changes are tracked.

There’s a level of discipline that comes from years of focus on user identity. However, machine identities often fall outside of that structure. They’re created quickly to support applications or automation. They’re granted the permissions needed to function, sometimes more than necessary. And over time, those permissions persist. These overprovisioned accesses are rarely audited, reviewed, and more importantly rarely reduced.

That’s where the gap forms.

It becomes difficult to answer basic questions about access. Not because the information doesn’t exist, but because it hasn’t been organized or managed in a way that makes it usable.

Visibility Before Control

When organizations start to address this problem, the instinct is often to tighten controls.

  • Limit permissions
  • Restrict access
  • Apply new policies

But control without visibility doesn’t solve much.

If you don’t understand how identities are being used, the business context of it in terms of  where they connect, what they interact with, and how they move across systems, then any attempt to restrict them becomes reactive and could result in slowing down business operations.

That’s why visibility needs to come first.

Once you can see how machine identities behave, patterns start to emerge. You can begin to understand where access is excessive, where dependencies exist, and where risk is concentrated. From there, governance can become more precise and more effective.

A Different Kind of Privilege Problem

Privilege sprawl isn’t new. Most organizations have spent years trying to manage excessive access among human users.

Machine identities introduce a similar issue, but with a different dynamic. Their access is often embedded into systems. It’s persistent, automated, and rarely questioned once it’s in place. That makes it harder to detect and easier to overlook.

And when something goes wrong, those identities can become a pathway for malicious actors to exploit

Where to Begin

For most organizations, the challenge isn’t awareness, it’s knowing where to start.

The first step isn’t a major transformation. It’s building clarity. Understanding how many machine identities exist. Where they’re being created. What permissions they have. How they’re used. And most importantly, confirming that a human user is mapped to a collection of non-human identities for the purposes of auditability and accountability.

Those questions sound simple, but they’re often difficult to answer. And that’s exactly why they matter. Because once you can answer them, you’re no longer operating in the dark.

Watch the Full Episode

In this installment of STRIVE, we go deeper into how machine identities are changing the way organizations should think about access, governance, and resilience. It’s a practical conversation about what’s happening now – and what needs to change moving forward.

Watch now.

Resource

If you’re interested in learning more about this topic, check out this e-book on non-human identities.

FAQs

Q: What is a machine identity?

A: A machine identity is a non-human identity used by applications, services, or systems to authenticate and interact with other resources.

Q: Why are machine identities becoming a bigger risk?

A: Because they are increasing in number, often have persistent access, and are not always governed as strictly as human users.

Q: How are they different from user identities?

A: They operate continuously, are embedded in automated workflows, and often lack structured lifecycle management.

Q: What is the biggest challenge organizations face in governing non-human identities?

A: Visibility. Many teams don’t have a clear understanding of how many machine identities are created, used, or interconnected.

Q: How does this impact resilience?

A: If compromised, machine identities can enable a malicious actor’s rapid movement across systems, making incidents harder to contain and recover from.

Q: Where should organizations start?

A: By identifying machine identities, understanding their permissions, and building governance practices that match their scale and complexity. And most importantly, confirming that a human user is mapped to a collection of non-human identities for the purposes of auditability and accountability

Vidya Shankaran is Field CTO at Commvault.

More related posts


Thumbnail_Blog_Identity-Resilience-Vishing_2026

Are You Ready for the Industrialized Vishing Attack?

Read more about Are You Ready for the Industrialized Vishing Attack?
Thumbnail_Blog-Identity-Resilience-MachineID-2026-Linkedin

The Machine Identity Blind Spot Is Now a Primary Attack Surface

Read more about The Machine Identity Blind Spot Is Now a Primary Attack Surface
Thumbnail_Blog-Help-Desk-2026-Linkedin

When the Help Desk Becomes the Front Door to Your Entire Network

Read more about When the Help Desk Becomes the Front Door to Your Entire Network

For decades, IT operations focused on uptime:

  • Keep the infrastructure running.
  • Hit your recovery time objective (RTO).
  • Meet your recovery point objective (RPO).

But modern cyber threats don’t respect infrastructure boundaries – and recovery isn’t just about restoring systems anymore. It’s about restoring clean, trusted data – across teams, under pressure.

In this episode of STRIVE, I sat down with Stephen Foskett, founder and president of the Futurum Group’s Tech Field Day, to discuss an emerging discipline: resilience operations – or ResOps.

And it’s more than a buzzword. It’s a shift in how organizations think about recovery intelligence.

Watch the full episode.

Key Takeaways: What ResOps Changes

  • ResOps moves recovery from infrastructure-focused to business-focused. It’s not just about bringing systems back online – it’s about restoring trusted, usable data.
  • Traditional RTO and RPO metrics aren’t enough anymore. Mean Time to Clean Recovery (MTCR) is emerging as a more meaningful way to measure resilience.
  • Breaking down silos is foundational to cyber readiness. Security, infrastructure, and DevOps must operate in sync – not in parallel.
  • Resilience is an operational discipline, not a tool. Culture, communication, and coordination matter as much as technology.
  • Recovery intelligence is becoming a competitive differentiator. Organizations that recover cleanly and quickly protect revenue, reputation, and trust. 

From IT Ops to ResOps: What’s Changed?

Stephen reflects on an earlier era of IT where teams often supported systems without fully understanding the business applications they powered. Recovery meant restoring infrastructure. Today, that model falls short. Modern environments are:

  • Distributed
  • Cloud-based
  • DevOps-driven
  • Security-sensitive
  • Deeply integrated with revenue streams

ResOps acknowledges that recovery is no longer an isolated IT function. It’s a cross-functional discipline that helps connect infrastructure, software development, and security with real business outcomes.

Why Traditional Metrics Don’t Tell the Whole Story

RTO. RPO. These metrics have guided disaster recovery planning for years. But as Stephen explains, restoring quickly isn’t enough if the data you restore isn’t clean.

Enter a more meaningful metric: MTCR. It’s not just how fast you recover; it’s how fast you can recover to a verified, trusted state.

In a ransomware event, that difference matters enormously. Restoring compromised data can restart an attack cycle. ResOps focuses on restoring operational integrity – not just functionality.

Sneak Peek: Why Clean Recovery Matters

In this moment from STRIVE, Stephen explains why traditional recovery metrics miss the mark – and why recovery is a cross-functional discipline.

The Real Barrier: Organizational Silos

Technology isn’t usually the biggest blocker to resilience. Structure is. Security teams often report to one executive. Infrastructure teams to another. Application teams to yet another. Each with different priorities, different incentives, and different definitions of success.

ResOps challenges that fragmentation.

Stephen discusses how collaborative workshops and cross-functional alignment are helping break down those silos. Because during a cyber event, organizational misalignment slows recovery more than tooling gaps ever will.

Why Commvault Is Leaning Into This Conversation

STRIVE isn’t about product features. It’s about how recovery thinking is evolving. ResOps aligns closely with what we see in the field:

  • Customers struggling with coordination during incidents.
  • Organizations restoring infrastructure but questioning data integrity.
  • Leadership asking for metrics that reflect real business impact.

The concept of MTCR reframes recovery intelligence around business trust – and that’s where the industry is heading. Recovery is no longer a back-office process. It’s an executive concern.

The Future of Recovery Intelligence

Looking ahead, ResOps is likely to mature rapidly. Over the next 12–18 months, organizations are expected to:

  • Integrate security and recovery workflows more tightly.
  • Adopt new recovery-focused metrics.
  • Operationalize resilience earlier in application lifecycles.
  • Invest in intelligence that distinguishes clean data from compromised data.

Cyber threats are accelerating. Recovery strategies must evolve at the same pace. ResOps helps provide a framework for doing that.

Watch the Full Episode

In this episode, we discuss:

  • How ResOps differs from traditional IT operations.
  • Why MTCR is helping to reshape recovery metrics.
  • What organizational alignment looks like in practice.
  • How DevOps culture influences resilience.
  • Where recovery intelligence is expected to head next.

Watch now.

If you’re responsible for cyber readiness, continuity, or recovery strategy, this is a must-watch discussion.

FAQs

Q: What is ResOps?

A: ResOps (Resilience Operations) is an emerging discipline that integrates IT operations, security, DevOps, and business stakeholders to help improve recovery intelligence and organizational resilience.

Q: How is ResOps different from traditional IT operations?

A: Traditional IT ops focuses primarily on infrastructure uptime. ResOps expands that focus to include clean data recovery, cross-functional coordination, and business alignment.

Q: What is Mean Time to Clean Recovery (MTCR)?

A: MTCR measures how quickly an organization can restore verified, clean data and resume safe operations after a cyber event – not just how quickly systems are brought back online.

Q: Why are metrics like RTO and RPO insufficient in modern environments?

A: They measure speed and data currency, but not data integrity. In ransomware scenarios, restoring compromised data can extend disruption.

Q: How can organizations start implementing ResOps?

A: Begin by:

    • Aligning security, infrastructure, and DevOps teams.
    • Evaluating recovery metrics beyond RTO/RPO.
    • Testing clean recovery processes.
    • Breaking down operational silos.
    • Incorporating resilience thinking earlier in system design.

Q: Why is recovery intelligence becoming more important?

A: As cyber threats grow more sophisticated, the ability to recover cleanly, quickly, and confidently directly impacts revenue, customer trust, and regulatory posture.

Darren Thomson is a Field CTO at Commvault.

More related posts


Thumbnail_Blog-Testing-Once-a-Year-2026

Testing Once a Year Is Not a Resilience Strategy

Read more about Testing Once a Year Is Not a Resilience Strategy
Thumbnail_Blog-IDC-Resops-2026

From Recovery to ResOps™: Building Enterprise Resilience That Scales

Read more about From Recovery to ResOps™: Building Enterprise Resilience That Scales
Readiverse-Featured-Image-888-x-500

Ready Is Good. Resilient Is Better.

Read more about Ready Is Good. Resilient Is Better.

Key Takeaways

  • Frontier AI is collapsing vulnerability remediation windows, prevention alone can no longer guarantee security.
  • The question that boards, regulators, and insurers are now asking is not “Do we have backups?” but “Can we prove we can recover cleanly?”
  • Backups are not recovery: a copy tells you data exists, not whether it is clean or restorable.
  • Mean Time to Clean Recovery (MTCR) must become a board-level, continuously measured number – not a theoretical estimate.
  • An Isolated Recovery Environment – air-gapped, immutable, hardened, and identity-isolated – is the baseline, not an advanced capability.
  • What counts as “clean” will keep changing as AI models grow more capable of finding compromises humans cannot anticipate.

I have spent a large part of my career running production systems. I know backup environments from the inside, the ones customers actually trust. I know recovery plans as the things that only reveal their weaknesses when something has already gone wrong. That experience changes how you think about cyber resilience.

From a distance, backup and recovery sounds manageable. Protect the data, store copies, document the runbook, test when you can, restore when you need to. But anyone who has run these environments at scale knows the harder truth: Recovery is where assumptions go to be tested. And right now, too many organizations are operating on assumptions that no longer fit.

For years, security operated on a familiar sequence: Find the vulnerability, patch it, harden the environment, monitor for activity. That model still matters. But the window it depends on is collapsing.

Frontier AI has changed the velocity of vulnerability discovery, attack path chaining, and exploit generation. Models like Claude Mythos and GPT-5.5-Cyber have already demonstrated what this looks like, so far in controlled, early-access testing that still relied on human expertise and carried meaningful false-positive rates, but the trajectory is unmistakable. As access widens, the same capability moves into attackers’ hands.

In a single month, Palo Alto Networks disclosed 26 CVEs, representing 75 underlying issues, after adopting frontier AI models for code scanning, compared with its typical volume of fewer than five CVEs per month.

Researchers also are warning that AI-assisted discovery is collapsing remediation windows, with some exploits now emerging within minutes of disclosure. When the patch window disappears, the remediation math stops working. Prevention cannot carry the full weight of readiness.

Prevention still matters, but it no longer defines readiness. The customers I talk to are not asking whether they need more controls. They already know they do. They are asking whether their business can recover cleanly when those controls fail, when attackers move faster than remediation cycles, or when compromise has been present longer than anyone realized.

That question is now what boards, regulators, and insurers are forcing. They have moved past “Do we have backups?” and toward something more consequential: “Can we prove we can recover cleanly?”

That proof starts with one distinction most organizations still get wrong: Backups are not recovery.

A backup tells you a copy exists. It does not tell you whether the data is clean, whether application dependencies are intact, whether identity services can be safely restored, or whether the recovery sequence still reflects the current environment.

I have reviewed plans that looked complete until someone tried to execute them. The runbook was there, but outdated. The restore worked but took three times longer than the estimate. The system came back, but downstream applications could not connect. None of that is unusual. It is exactly what real testing is supposed to surface. The problem is most organizations discover these gaps during an actual incident.

The metric that matters most when something goes wrong is how quickly you can return to a known-good state. That is why Mean Time to Clean Recovery (MTCR) needs to become a board-level number, not a theoretical estimate in a plan, but a measured, validated time.

The Moving Target: What’s Clean Today May Not Be Clean Tomorrow

With Frontier AI models, the honest answer is this: you cannot guarantee that every vulnerability will be found and remediated in time. Attackers leveraging the same models are discovering and chaining exploits faster than any remediation program can realistically keep pace with. That is not a failure of your security team. It is the new physics of the threat landscape.

What you can control is your ability to recover. That means an Isolated Recovery Environment – backups air-gapped from the internet, unreachable from the production network, and protected from the lateral movement that defines a sophisticated breach. It means immutability and compliance lock, so no credential, however privileged, can shorten retention or delete data outside an authorized process. And it means ResOps in practice: not just backing up data, but continuously testing recovery, automating integrity validation, and measuring your MTCR – the validated time to return to a known-good state.

But here is the part most organizations are not yet accounting for: what counts as “clean” is not a fixed line. As AI models grow more capable, they will increasingly find vulnerabilities that the human mind simply cannot anticipate, novel attack paths, dormant implants, subtle corruptions embedded long before detection. A recovery point that is clean by today’s standards may carry compromise that tomorrow’s AI-assisted forensics will surface. That means your definition of clean must evolve continuously. MTCR is not a number you set once. It is a discipline you maintain, revisiting what clean means, updating your validation criteria, and treating resilience as a living standard rather than a certification you pass once.

So what is a good MTCR? Based on what I have seen work in practice, the target for your entire minimum viable company – the smallest set of systems that lets you keep operating, which I define precisely below, should be under six hours. Six hours is achievable with the right architecture: an IRE ready to run, a pre-validated recovery sequence, and runbooks that are executable rather than readable. If your current MTCR is measured in days, the gap is almost always one of those three.

Four Steps to Stay Resilient in the Frontier AI Era

Accepting that prevention alone is not enough is the starting point. From there, the work gets specific. Here is where I tell organizations to focus.

1. Evaluate your actual recovery risks.

Most recovery risk assessments ask the wrong questions. “Do backups exist?” is not the same as “Can we recover cleanly?” The harder questions are: Can critical systems be restored without reintroducing the threat? Are recovery environments isolated from compromised production systems? Are recovery plans mapped to current dependencies – not the architecture from two years ago?

In a fast-moving vulnerability environment, the gap between “we have backups” and “we can recover” is where organizations get hurt. Assessing that gap honestly, before an incident forces the issue, is where resilience planning must start. That assessment needs to include a business impact analysis: which systems have a recovery window measured in minutes, which in hours, and which can wait a day. Without that tiering, every system looks equally urgent during an incident, and nothing gets restored fast enough.

2. Make isolated recovery and air gapping the baseline – not the exception.

If you are still treating air-gapped, immutable copies as an advanced capability rather than a standard requirement, that assumption no longer holds. When exploitation timelines compress to minutes, you need fallback options that are structurally separated from production identity, network, and management planes – logically or physically isolated, immutable, and with no live path back to production that an attacker can follow.

The goal is not just protection from the current threat, but maintaining clean recovery options when a vulnerability you have not patched yet gets exploited. That happens now. Plan for it.

Isolation only holds if the infrastructure around it is hardened. That means backup infrastructure on hardened operating systems, not generic images, and ideally on physical servers that survive a hypervisor-layer attack. It means encryption keys stored outside the backup platform, in an external vault with just-in-time access and no dependency on production Active Directory. And it means treating your backup domain as a separate identity boundary: no trust to production AD, mandatory MFA, and multi-person authorization for destructive operations. None of this is exotic, it is the baseline for your environment to recover into an uncompromised space.

Equally important is the question of what you are recovering from. Industry incident-response data consistently puts median breach dwell time in the range of weeks, not days. That means your recovery copies need to reach back far enough to find a genuinely clean point, not just yesterday’s backup. Critical systems warrant multiple geographically separated copies, including at least one immutable copy and one that is fully offline. Retention policy is not a storage cost decision. It is a security decision.

3. Know which systems the business cannot operate without – and recover those first.

Most organizations discover their recovery sequence during an incident. That’s why the first 24–48 hours aren’t spent restoring systems, they’re spent deciding what matters.

Organizations know they have to recover identity platforms, billing systems, operational databases, and core infrastructure. What they often have not mapped is the order, the dependencies between those systems, and the downstream applications that cannot function until specific services are back.

This gets more complex as AI becomes embedded in business operations. Data pipelines, model repositories, vector databases, agentic workflows – these are now operational dependencies, not just technical infrastructure. If your recovery sequencing does not account for them, your recovery time estimates are probably wrong.

Defining what it means to operate as a minimum viable company (the smallest set of systems required to keep the business running) and building recovery around that definition is not a theoretical exercise. It is the practical answer to the question every executive team will ask during an incident: What do we bring back first?

In my experience helping customers through active incidents, the first 12 hours answer that question whether you have planned for it or not – what gets recovered in that window becomes your MVC by default. The organizations that come through fastest decided in advance: they knew exactly which systems had to be back within 12 hours and had validated they could do it. If your MVC does not fit in 12 hours, it is not your MVC, it is a wish list. The work is to keep trimming until what remains can realistically be restored in that window, then test it until you can prove it.

4. Automate resilience and test continuously – not on a calendar schedule.

A recovery plan that lives in a document and is reviewed annually is not a recovery capability. It is a hypothesis that has never been tested against reality.

The problem with calendar-based testing is what it misses between cycles. Environments change constantly: new workloads, updated dependencies, infrastructure that has drifted from what the runbook describes. By the time the annual test runs, it is validating a snapshot of an environment that no longer exists. In a threat landscape where exploitation can happen within minutes of disclosure, that lag is not acceptable.

Threat scanning, clean recovery point identification, dependency-aware restoration, and recovery orchestration all need to be automated and running continuously. Not because automation is a best practice, but because the manual alternative cannot keep pace with how fast things now move.

Continuous testing also depends on continuous detection. You cannot select a clean recovery point if you do not know when the compromise began. That is why threat detection, anomaly scanning of backup data, and recovery-point analysis have to feed each other: detection tells you which copies predate the intrusion, and that determination drives which point you actually recover from. Without that link, you are restoring to a date you hope is clean rather than one you have verified, and in a Frontier AI threat landscape, hope is not a recovery strategy.

What continuous testing surfaces is different from what annual tests find. Calendar tests tend to confirm the plan works under controlled conditions. Continuous testing finds the dependency that changed last month, the recovery sequence that breaks when a specific workload is added, the identity service that takes twice as long to restore as the estimate assumed.

Those are the gaps that matter during a real event, and the only way to find them before an incident does is to be testing all the time.

Testing also needs to happen in the right environment. A recovery test that runs against production infrastructure does not tell you whether you can recover when production is compromised. Cleanroom testing – validating restoration in a fully isolated environment with no connectivity back to production – is how you confirm your backup copies are genuinely usable under incident conditions. That includes recovering identity services, external key management, and Tier 0 applications in isolation, with dedicated break-glass accounts that exist outside your normal directory.

What makes daily testing viable is validate restore, a recovery type that exercises the full restore path for every critical asset without touching production. Your backup platform needs to support this natively; if it cannot run an automated, non-disruptive recoverability test across your MVC every day, you do not actually know whether your backups work. In Commvault, this restores against your critical-asset groups, with automated reporting on the recovery status of every protected system.

The same applies to your runbooks. A runbook that lives in a Word document or PDF is a reference manual, not an operational tool – it assumes someone has the time, clarity, and access to read it under pressure. Real runbooks are digital scripts that execute the recovery sequence and validate each step, confirming the application actually works before moving on: not the service started” but “the application responded correctly to a synthetic transaction.” Commvault’s Cleanroom Runbooks are built for this – executable workflows that drive an end-to-end recovery in an isolated environment without a human interpreting a document at every step.

One final point that rarely makes it into recovery plans until it is too late: during a serious incident, your corporate communications infrastructure may itself be compromised or unavailable. Email, Teams, and Slack run on the same infrastructure attackers target. Know in advance which out-of-band channels your team will use to coordinate, and make sure those channels are tested alongside your technical recovery procedures.

Hear more from Commvault’s Chief Security Officer Bill O’Connell on the four critical steps for resiliency in the AI era.

Resilience Is an Operating Discipline, Not a Project

The organizations that will hold up under frontier AI-accelerated threats are the ones that treat resilience as an operating discipline — measured MTCR, continuous validation, and a recovery capability they have proven, not assumed.

The problem isn’t that attacks are getting faster. It’s that recovery hasn’t caught up, and until it does, the math doesn’t work.


FAQs

Q: What is Mean Time to Clean Recovery (MTCR) and why does it matter?
A: MTCR measures how quickly an organization can return to a verified, known-good state after a cyberattack – not just restore data, but confirm it is clean and that application dependencies are intact. It should be a board-level metric with a measured, validated time, not a theoretical estimate buried in a recovery plan. The target for a well-architected MVC – covering all identity systems, critical applications, and isolated environment readiness – is under six hours.

Q: What is an Isolated Recovery Environment and how is it different from a standard backup?
A: An Isolated Recovery Environment is a fully air-gapped, immutable copy of critical data that is structurally separated from production networks, identity systems, and management planes. A standard backup tells you a copy exists. An IRE tells you that copy is protected from the same attack that hit your production environment.

Q: How do we know whether we can actually recover today?
A: The only honest answer comes from testing, not documentation. If you cannot point to a recent, validated recovery of your minimum viable company – ideally a daily automated test – then you do not know, you are assuming. A defensible answer to the board is a measured MTCR backed by continuous validation, not a recovery plan that looks complete on paper.

Q: What do regulators and cyber insurers now expect?
A: The bar has moved from “Do you have backups?” to “Can you prove you can recover cleanly, and how fast?” Regulators increasingly expect demonstrable recovery capability and tested resilience; insurers increasingly price coverage – and pay claims – based on evidence of isolated, immutable backups and validated recovery times. A measured MTCR and a documented testing cadence are becoming table stakes for both.

Rajiv Kottomtharayil is Chief Product Officer at Commvault.

More related posts


Thumbnail_Blog-What-is-Resops-2026

What Is ResOps – and Why Cyber Resilience Needs It

Read more about What Is ResOps – and Why Cyber Resilience Needs It
Thumbnail_Blog-Playbook-2026

Prove It Before You Need It: The Playbook I Wish I’d Had 10 Years Ago

Read more about Prove It Before You Need It: The Playbook I Wish I’d Had 10 Years Ago
Thumbnail_Blog-Medusa-is-Evolving-2026

Medusa Is Evolving. Cyber Resilience, Cyber Recovery, and ResOps Matter More Than Ever.

Read more about Medusa Is Evolving. Cyber Resilience, Cyber Recovery, and ResOps Matter More Than Ever.

Blog

Protecting AI Workloads: How Can Organizations Achieve Resilience in the AI Era?

AI resilience helps enable protection, recovery, and governance of AI workloads, data, and models by combining threat detection, clean recovery, and controlled data access.

Frequently Asked Questions

What is AI resilience?

AI resilience is the ability to protect, recover, and govern AI systems across their full lifecycle. Commvault’s Protect and Leverage AI capabilities help verify that data, models, and pipelines remain secure, recoverable, and trustworthy — even when disrupted by cyber threats, failures, or operational complexity in hybrid and multi-cloud environments.

Why is protecting AI workloads important?

AI workloads rely on distributed data, models, and infrastructure, making them vulnerable to threats like data poisoning and model corruption. Protecting them helps uphold data integrity, reduce operational risk, and maintain trust in AI-enabled business processes. Commvault helps address these challenges with Metallic AI, unifying ML-driven detection, guided recovery, and automation across Commvault Cloud.

What does full AI stack protection include?

Full AI stack protection safeguards data pipelines, vector databases, models, metadata, configurations, and compute infrastructure. Commvault Cloud Unity covers this breadth — including unified data platforms like Amazon Redshift and Google BigQuery, vector retrieval systems, and compute infrastructure — enabling complete and consistent recovery of AI workloads across hybrid and multi-cloud environments.

Why does clean recovery matter in AI environments?

Clean recovery confirms that restored data is free from corruption, malware, or inconsistencies. In AI systems, compromised data leads to inaccurate outputs and biased decisions. Commvault Synthetic Recovery addresses this by analyzing multiple backup versions to assemble a validated recovery point — so restored AI workloads produce trusted, accurate outputs.

How does AI improve data protection and operations?

Commvault embeds AI across the protection lifecycle — automating threat detection, optimizing backup scheduling, and predicting storage needs through ML-enabled capabilities. Arlie, Commvault’s AI assistant, improves user experience through natural language interactions, guided workflows, and intelligent insights, helping security and IT teams manage complex AI environments more efficiently.

What is responsible AI in data protection?

Responsible AI enables systems to operate with transparency, governance, and control. Commvault supports this through Data Activate — a governed workspace that applies encryption, immutability, and role-based access controls to curate and extend trusted data to AI and analytics platforms, helping prevent misuse and maintain compliance while enabling innovation.


Key Takeaways

  • Mythos will likely accelerate vulnerability discovery to a scale and speed that outpaces traditional, human-driven remediation workflows.
  • Core security fundamentals like patching, air-gapped backups, and disciplined vulnerability management remain critical but may no longer be sufficient on their own.
  • The primary challenge is shifting from detection to the capacity to act as vulnerability volumes surge beyond current operational limits.
  • AI resilience depends on the ability to recover coherent systems – not just data – across models, pipelines, and permissions.
  • Organizations that proactively adapt during this early window are likely to be significantly better positioned than those that delay action.

A few weeks ago, I was in a room with a group of CIOs and CISOs when the conversation turned to Mythos and Project Glasswing. The energy was immediate – these are people who have lived through a lot of hype cycles, and this commanded their attention.

The reactions landed in two camps. One: The threat categories aren’t new – organizations with solid vulnerability management and trusted air-gapped backups will be better positioned than those without. Two: The velocity is different – not just what Mythos can find, but how fast, how fast bad actors could leverage AI for machine-speed attacks, and what that does to the math that most vulnerability management programs are built on.

Both were right. That’s what made the conversation worth writing about.

What Mythos Changes – And What It Doesn’t

Mythos is Anthropic’s AI model for autonomous vulnerability discovery. It can find and chain critical exploits across major operating systems at a success rate that is believed to have no real precedent in this domain.

Project Glasswing – the consortium of companies brought in to test and harden their systems before Mythos or similar capabilities reach adversaries – is the signal that this is real, it is here, and the window for getting ahead of it is short.

The fundamentals-first view holds: patching matters, virtually air-gapped backups matter, vulnerability management discipline matters. None of that changes with Mythos. What changes is the production rate on the other side of those programs.

The question after Glasswing isn’t whether you have a vulnerability management program. It’s whether it was built for findings that arrive in a trickle – or a tsunami.

Most programs were built for the trickle. Periodic assessments, CVSS-based prioritization queues, patch and testing cycles measured in weeks. That cadence made sense when the pace of discovery matched the pace of human-led processes. Mythos-class capability breaks that assumption – the volume of exploitable findings may exceed what most organizations can process through the workflows they have today.

The issue isn’t detection. It’s capacity to act – and what happens when the gap between discovery and remediation widens faster than you can close it.

When Prevention Gets Compressed, Resilience Moves Forward

When prevention timelines are compressed, the resilience question moves to the front of the line. If you can’t guarantee you’ll patch everything before something is exploited – and increasingly, you can’t – the questions that matter shift: How fast do you detect? How do you contain? And when you recover, what exactly are you recovering to?

That last question is harder than it sounds, especially for organizations with progressive agentic interactions. An AI system isn’t just data. It’s a model version, a training pipeline, a vector database, a set of agent identities and permissions – all of which need to reflect the same operational state to constitute something you can actually trust.

Most organizations can restore individual components. Very few can prove that what they’ve restored is coherent.

Recovering an AI system isn’t a data restoration problem. It’s a coherence problem – and the gap between those two things is where most enterprises are currently exposed.

This is the thread that connects Mythos to the broader AI resilience conversation. It isn’t that Mythos introduces a new type of risk that requires a new framework.

It’s that Mythos compresses the timeline in a way that surfaces existing gaps faster, with less runway to close them before something goes wrong, thereby increasing the change that something will go wrong before an organization can properly remediate vulnerabilities.

The Window Is Open. It Won’t Stay That Way.

Glasswing was designed to give defenders a head start. The organizations that use this window deliberately – stress-testing their vulnerability programs for volume, getting AI resilience infrastructure to a state they can defend, and treating recovery as something that has to be provable before an incident, not assembled during one – will be in a materially better position than those that wait.

The fundamentals still apply. The urgency is new.

The Agentic Enterprise: Why AI Resilience Demands a System of Record – Commvault’s latest Readiness Report – examines the AI resilience infrastructure gaps that determine whether organizations can answer the hard recovery questions when the pace of threats demands it.

FAQs

Q: What is Mythos and why is it significant?

A: Mythos is an AI model designed for autonomous vulnerability discovery, capable of identifying and chaining exploits across systems at unprecedented speed. Its significance lies in how it compresses the timeline between vulnerability discovery and potential exploitation, raising the stakes for defenders.

Q: Does Mythos change the fundamentals of cybersecurity?

A: No, core practices like patching, backups, and vulnerability management still matter. What is changing is the volume and velocity of threats, which puts pressure on existing processes that were designed for slower, more predictable workflows.

Q: Why may current vulnerability management programs struggle?

A: Many programs were built for a steady flow of findings, not the surge enabled by AI-driven discovery. As a result, organizations face a growing gap between identifying vulnerabilities and actually remediating them.

Q: What does “resilience” mean in the context of AI systems?

A: Resilience goes beyond restoring data – it involves recovering an entire AI system in a coherent, trustworthy state. This includes models, training pipelines, vector databases, and access controls all aligning correctly.

Q: Why is recovery becoming more important than prevention?

A: As prevention timelines shrink due to faster exploitation, it is becoming unrealistic to patch everything in time. This shifts focus to how quickly organizations can detect, contain, and recover from incidents.

Q: How can organizations start preparing?

A: Organizations can stress-test their vulnerability management processes, modernize resilience infrastructure, and validate recovery capabilities. Acting during this early window provides a meaningful strategic advantage.

Tim Zonca is Vice President, Portfolio Management, at Commvault.

More related posts


Thumbnail_Blog-Anthropic-Project-ResOps-2026

Anthropic’s Project Glasswing Makes the Case for ResOps

Read more about Anthropic’s Project Glasswing Makes the Case for ResOps

Key Takeaways

  • Agentic AI introduces new security risks because it plans, remembers, and acts across systems instead of stopping after a single prompt-response cycle.
  • Poisoned training data can quietly influence model behavior at scale, even when the model still appears to perform normally in standard testing.
  • Compromised vector databases can steer agent decisions by corrupting the context the model relies on, making bad behavior look legitimate.
  • Ungoverned agent identity creates a machine-speed access-control problem that traditional human-centric identity systems are not built to handle.
  • Cascading decisions built on bad state can spread corruption across multiple agents and workflows, making rollback and recovery much harder.

The tools, controls, and governance policies most enterprises have in place were designed for systems that answer questions – retrieval tools, copilots, generative assistants. Systems that respond to a prompt and stop. When something went wrong, the failure was discrete. Fix the prompt, adjust the configuration, move on.

Agentic AI doesn’t work that way. These systems plan, remember, and execute across the enterprise without step-by-step human instruction. They maintain state. They coordinate with other agents. They act on production systems – writing to databases, triggering workflows, making decisions at machine speed.

That architectural shift introduces four threat vectors that existing security frameworks were never designed to address. If your AI governance strategy doesn’t account for them, you likely have exposure you probably can’t see.

1. Poisoned Training Data

An AI system is only as trustworthy as the data it was trained on. That statement has always been true. What’s changed is the attack surface.

In agentic AI deployments, training pipelines are larger, more complex, and frequently assembled from multiple sources – internal data, third-party feeds, vendor-provided datasets. Each dependency in that chain is a potential injection point. An adversarial actor who can influence training data – through supply chain compromise, insider access, or contamination of a shared data source – can shape model behavior at scale.

What makes this particularly dangerous is that poisoned models often perform normally on standard benchmarks. The manipulation may be surgical: designed to produce specific outputs in specific contexts while behaving correctly everywhere else.

By the time the effect surfaces in production, the model has been in use for weeks or months, and tracing the contamination back to its source requires exactly the kind of relational data provenance most organizations don’t have.

The question to ask: Can you produce a complete, verifiable record of what data your models were trained on – at a specific point in time?

2. Compromised Vector Databases

Vector databases are the memory layer of agentic systems. Before an agent acts, it queries a vector store to retrieve relevant context – past interactions, domain knowledge, reference data – that shapes what it does next.

Most security teams aren’t thinking about vector databases the way they think about other sensitive data stores. They should be.

A compromised vector database doesn’t just return wrong answers. It shapes the decisions that follow. Injected embeddings – malicious content inserted into the vector store – can redirect agent behavior in ways that appear completely legitimate from the outside.

An agent asked to approve a transaction retrieves context that subtly reframes the approval criteria. An agent managing customer communications pulls context that steers responses in an attacker’s preferred direction. The action looks correct. The reasoning looks sound. But the underlying context has been manipulated.

This attack vector is particularly hard to detect because it operates below the model layer. Standard model monitoring won’t catch it. The model is behaving exactly as trained – it’s the context it’s reasoning from that’s been corrupted.

The question to ask: Is your vector database treated as a sensitive, governed data asset – with access controls, integrity monitoring, and audit logging comparable to your most critical production databases?

3. Ungoverned Agent Identity

In a multi-agent architecture, agents don’t just interact with data – they interact with each other. They spawn subagents, delegate tasks, request outputs, and synthesize results from agents they’ve never been explicitly connected to. To do this, they authenticate, present credentials, and establish trust.

Agent identity is the access control layer for the autonomous enterprise – and it’s a gap that identity security vendors and identity providers (IDPs) don’t close. Their governance frameworks are built for human identity.

Agent identities created within those same rules appear completely legitimate: They were provisioned correctly, they followed policy. The IDP isn’t failing – it simply has no framework for determining whether an agent is acting outside the context it was created for, has been quietly escalated, or is coordinating where it shouldn’t be.

The exposure is qualitatively different from traditional credential compromise. When a human user’s credentials are stolen, the attacker operates within that user’s permissions, at human speed.

When an agent’s identity is compromised, the attacker gains access to the autonomous decision-making layer – the ability to trigger workflows, approve actions, coordinate with other agents, and exfiltrate data at machine speed, at scale, through channels that appear entirely normal.

Identity-layer failures are also among the hardest to detect after the fact. Agent actions taken under a compromised identity don’t look anomalous – they look like legitimate agent behavior. And because they’re generated by a system rather than a human, the volume can be enormous before anyone notices.

Recovery compounds the problem. Most AI recovery playbooks focus on restoring data: training sets, model weights, pipeline configurations. Identity is rarely on the list. A system recovered with clean data but misaligned identity configurations isn’t actually recovered. It’s a clean system with a poisoned access layer.

The question to ask: Is agent identity managed with the same rigor as human identity – with lifecycle management, least-privilege access, and inclusion in recovery playbooks?

4. Cascading Decisions Built on Bad State

The first three attack vectors are discrete. This one is systemic – and in many ways it can be the most difficult to contain.

Multi-agent architectures are designed for coordination. Agents share context, pass outputs to one another, and build on each other’s work. That coordination is what makes them powerful. It’s also what makes failures propagate.

An agent operating on corrupted memory doesn’t fail cleanly. It produces outputs – decisions, actions, data – that other agents consume. Those agents produce their own outputs. By the time the original corruption surfaces as something observable, bad state may have touched dozens of downstream processes, across multiple agents, with no clean rollback path.

This is what makes the context gap so significant. At any given moment, your AI system consists of a model version, a set of training data, an artifact store, a pipeline configuration, and a set of active agent interactions – all of which need to reflect the same operational state to constitute a trustworthy, recoverable system. When they don’t, you don’t just have an error. You have a system that is coherent in pieces and incoherent as a whole.

Point tools can each confirm their own slice. None can confirm the pieces belong together. That’s not a monitoring problem you can solve by adding another tool. It’s a structural gap – and the only way to close it is with a system that captures AI state relationally: what was running, against what data, with what configuration, at what moment.

The question to ask: If your AI infrastructure were compromised today, could you identify exactly what state every component was in before the incident – and prove it?

What This Means for Your Security Strategy

Each of these four vectors requires a different defensive response. But they share a common implication: The governance and resilience frameworks designed for the previous era of AI don’t cover the failure modes of the agentic era.

Securing agentic AI requires extending your framework in three directions:

  • Deeper, into the data and identity layers that sit below the model.
  • Broader, to cover agent-to-agent interactions that existing monitoring doesn’t observe.
  • Relationally, to capture not just the state of individual components, but how they fit together at any point in time.

That last requirement is the one most organizations haven’t yet confronted. And it’s the one that will determine whether, when something goes wrong, you have a recoverable system or a collection of accurate-looking reports describing something that no longer exists.

Read The Agentic Blind Spot: Why AI Resilience Demands a System of Record to learn why you need an SOR to help protect the consistency and accuracy of your AI data.

FAQs

Q: Why are agentic AI systems riskier than traditional generative AI tools?

A: Agentic systems do more than answer prompts. They maintain state, coordinate with other agents, and take actions in production environments, which expands the attack surface far beyond simple prompt manipulation.

Q: What makes poisoned training data so difficult to detect?

A: The manipulation can be highly targeted, affecting only specific situations while leaving normal benchmarks intact. That means a model may look healthy until the poisoned behavior appears in real use.

Q: How can a vector database become a security problem?

A: A vector database shapes the context an agent uses before acting. If that context is altered, the agent may make decisions that seem reasonable on the surface but are really being guided by malicious data.

Q: Why is agent identity different from human identity?

A: Agent identity is tied to autonomous actions, delegation, and machine-speed execution. Traditional identity governance is designed for people, so it often misses whether an agent is acting outside its intended context.

Q: Why is cascading bad state such a serious issue in multi-agent systems?

A: Once one agent consumes corrupted output, that error can spread to downstream agents and workflows. The result is not just one bad decision, but a chain of connected failures.

Q: How can organizations improve AI security?

A: Extend governance deeper into data and identity layers, monitor agent-to-agent interactions, and track AI state relationally to enable them to reconstruct what happened during an incident.

Michael Thelander is Senior Director, Product Marketing, at Commvault.

Related Blogs

More related posts


Thumbnail_Blog-Data-Access-Governance-2026

Securing AI with Unified Data Access Governance

Read more about Securing AI with Unified Data Access Governance
Thumbnail_Blog-Environmental-Footprint-AI-2026

Smarter Data, Greener AI

Read more about Smarter Data, Greener AI
Thumbnail_Blog-Anthropic-Project-ResOps-2026

Anthropic’s Project Glasswing Makes the Case for ResOps

Read more about Anthropic’s Project Glasswing Makes the Case for ResOps
Thumbnail_Blog-Data-Rooms-2025-Linkedin

Data Activate: Unlocking the Power of Trusted Data for AI Innovation

Read more about Data Activate: Unlocking the Power of Trusted Data for AI Innovation
Thumbnail_Blog-AI-Agents-2026

AI Agents Are Everywhere. Do You Know What They’re Doing?

Read more about AI Agents Are Everywhere. Do You Know What They’re Doing?
Thumbnail_Blog-Building-AI-Agents-2026

From Experimentation to Operation: Building AI Agents You Can Actually Trust

Read more about From Experimentation to Operation: Building AI Agents You Can Actually Trust

Key Takeaways

  • Agentic AI systems are stateful and continuously operating, making traditional recovery models insufficient.
  • The memory layer (vector databases and context storage) is a critical yet under-monitored attack surface.
  • Runtime decision-making workflows can be manipulated without triggering traditional security alerts.
  • Observability gaps in agent-to-agent interactions leave most organizations with incomplete visibility into risks.
  • True recovery requires a unified, time-aligned record of all system layers to restore a trustworthy state.

Most enterprises entering the agentic AI era are managing resilience with the wrong mental model – and the data backs it up: Only 1 in 5 companies has a mature model for governing autonomous AI agents. They’re thinking about AI the way they think about applications: discrete, stateless, recoverable by restoring clean data to a clean environment.

Agentic AI doesn’t work that way. These systems are stateful, continuously operating, and architecturally layered in ways that create failure modes most security and resilience frameworks weren’t designed to address. The gap isn’t in tooling. It’s in understanding what’s actually running – and what “recovery” has to mean for systems built this way.

There are four architectural layers that define the problem. Each one is distinct. Each one is underprotected. And together, they explain why an agentic AI system can appear recoverable while remaining fundamentally compromised.

Layer 1: Agent Memory – The Attack Surface You’re Not Watching

Traditional enterprise applications don’t remember anything between sessions. Agentic AI does. The memory layer – primarily vector databases storing embeddings, but also session state and retrieved context – is what gives agents continuity across interactions. It’s what allows an agent to pick up where it left off, to draw on prior context, to build a coherent picture of a complex workflow over time.

It is also one of the most consequential attack surfaces in the modern enterprise stack – and one of the least monitored.

The attack vector is subtle enough to evade most conventional security tooling. An adversary who can influence what gets written to a vector database can shape what the agent believes to be true. Injected or manipulated embeddings don’t need to look malicious – they need to look authoritative.

A compromised memory store can redirect agent behavior, exfiltrate data through agent actions, or cause an agent to make decisions that appear legitimate but serve an attacker’s objectives. None of this requires touching the model itself.

The detection problem is compounded by the volume and velocity of vector database writes in active agentic deployments. Anomaly detection tools built for structured data don’t translate well to embedding space. The signal is there – but most organizations aren’t equipped to read it.

What resilience requires here: continuous integrity monitoring of vector databases, not just backup. Version-controlled embeddings with a provable chain of custody. The ability to identify, at any point in time, exactly what the memory layer contained – and to restore to a verified clean state, not just a recent one.

Layer 2: Runtime Control – When the Workflow Is the Threat

Agentic AI doesn’t execute fixed scripts. It plans. At runtime, an agent receives a goal, determines the steps required to achieve it, selects the tools it needs, and executes – often spawning subagents to handle parallel workstreams. The workflow is dynamic, constructed in the moment, and frequently long-running.

This is what makes agentic AI genuinely useful. It’s also what makes it genuinely difficult to protect.

In a conventional automation environment, a compromised workflow is bounded. It does what it was configured to do, and it stops. A compromised agentic workflow is different: It adapts.

If an attacker can influence the planning layer – through a poisoned prompt, a manipulated tool response, or a corrupted planning model – the agent will pursue the attacker’s objective using whatever legitimate tools and access it has. It will look like normal operation. The logs, to the extent they exist, will show authorized tool calls.

Consider a procurement agent tasked with validating vendor invoices against contract terms. Under normal operation, it checks invoice amounts, cross-references approval thresholds, and flags exceptions for human review.

An attacker who can influence the planning layer – through a manipulated tool response from the contract database – doesn’t need to touch the approval logic directly. They simply give the agent a contract record with altered thresholds.

The agent plans correctly against corrupted inputs. Every tool call it makes is legitimate. Every decision it reaches is wrong. By the time the anomaly surfaces in a finance reconciliation, the workflow has processed weeks of invoices and the audit trail shows nothing but authorized actions.

The window between compromise and detection in these scenarios is not measured in seconds. Agentic workflows operate continuously. By the time anomalous outcomes surface, the workflow may have touched dozens of systems, made hundreds of decisions, and left changes across production environments that are difficult to enumerate and harder to reverse.

What resilience requires here: runtime monitoring that watches what agents are deciding, not just what they’re doing. Intervention mechanisms that can halt a running workflow cleanly without cascading failures. Recovery playbooks built for long-running agentic processes – not just for discrete transactions.

Layer 3: Agentic Observability – The Logging Gap at Machine Speed

Enterprise logging infrastructure was built for human-scale operations. It captures what systems do, at a granularity and latency designed for human review. Agentic AI operates at a different speed entirely.

In an active multi-agent deployment, agents are spawning subagents, passing context between one another, making tool calls, and synthesizing outputs – continuously, in parallel, faster than conventional logging pipelines were designed to capture.

The interactions that matter most for security – agent-to-agent communications, context handoffs, tool invocations that cross trust boundaries – are exactly the interactions that existing monitoring frameworks leave most underobserved.

Today, only 17% of continuously monitor agent-to-agent interactions. The other 83% are governing agentic AI based on a partial picture – one that captures what individual agents do in isolation but misses the interaction layer where the most consequential security events occur.

This isn’t a gap that more logging volume solves. The problem isn’t the quantity of data being captured – it’s that the data structures and latency requirements of agentic interactions don’t fit well into observability frameworks designed for slower, more structured systems. Closing this gap requires purpose-built agentic observability tooling, or significant adaptation of existing infrastructure.

What resilience requires here: end-to-end visibility into agent-to-agent interactions, not just individual agent outputs. Logging architectures that can operate at agentic speed without dropping events. The ability to reconstruct, after the fact, the full sequence of agent decisions and interactions for any given workflow.

Layer 4: Multi-Agent Coordination – Where Emergent Failures Hide

The most architecturally novel risk in agentic AI doesn’t come from any single compromised agent. It comes from how agents depend on one another – and how failures propagate across those dependencies before anyone realizes something is wrong.

In a multi-agent architecture, agents share context. An orchestrator agent passes a task brief to a subagent; the subagent returns a result that the orchestrator incorporates into its next decision.

If the subagent’s output is corrupted – through a compromised memory layer, a manipulated tool response, or a poisoned planning model – the orchestrator has no native way to detect it. It treats the output as authoritative. It incorporates it. It acts on it. And it passes its own now-compromised output downstream.

This is the emergent failure mode: a corruption that originates in one layer, propagates through agent interactions, and surfaces as an anomalous outcome in a system several steps removed from the original compromise. By the time it’s visible, the causal chain is long and the blast radius is significant.

Consider a threat intelligence pipeline where a data-gathering agent ingests feeds from external sources, a classification agent categorizes and scores them, and an orchestrator incorporates the scored intelligence into security posture recommendations pushed to downstream teams.

If the data-gathering agent’s memory layer is compromised – subtly, through injected embeddings that cause it to weight certain threat actors as low-risk – the classification agent receives inputs it has no reason to question. It classifies accurately against what it’s given.

The orchestrator incorporates the results confidently. Security teams downstream deprioritize the relevant threat category based on what looks like a coherent, multi-source consensus. The failure originated in Layer 1. It expressed itself in Layer 4. Nothing in between flagged an anomaly because nothing in between had visibility across the full chain.

The governance frameworks most enterprises apply to AI were designed for model outputs – what the AI says. Multi-agent coordination failures are not model output failures. They are systems failures, arising from the interaction layer between models, and they require a different kind of governance: one that monitors and controls not just individual agent behavior but the trust relationships between agents, the integrity of context as it passes between them, and the access rights that govern what any agent can request of any other.

What resilience requires here: agent identity management that treats inter-agent trust as a first-class security concern. Integrity verification for context as it moves across agent boundaries. Governance policies that cover autonomous agent behavior – not just the outputs of individual models.

The Relational Problem That Ties All Four Together

These four layers are distinct in their failure modes, but they share a common vulnerability: none of them has a shared record of how they relate to each other at a specific point in time.

The model registry knows what version is running. The vector database knows what’s in memory. The orchestration layer knows what workflow is active. The identity system knows what agents have what access. Each can confirm its own slice of the picture. None can confirm whether those slices belong together – whether they reflect the same operational state, the same moment, the same trustworthy configuration.

That’s the context gap. And it’s why recovery from an agentic AI compromise isn’t a data restoration problem. It’s a coherence problem – one that requires a unified record of the relationships between layers, not just the components themselves.

Close this gap before an incident, or spend an incident trying to close it.

The architecture challenges covered here are only part of what security and resilience leaders need to understand about agentic AI risk. The Agentic Blind Spot: Why AI Resilience Demands a System of Record goes further, examining where most enterprises actually stand on AI resilience readiness, what the governance gaps look like in practice, and what it takes to make “our AI is trustworthy” a provable claim, not just an assertion.

FAQs

Q: Why doesn’t traditional disaster recovery work for agentic AI?

A: Traditional recovery assumes systems are stateless and can be restored from clean backups. Agentic AI systems retain memory, evolve over time, and depend on layered interactions, making simple restoration insufficient to regain trust.

Q: What makes the memory layer in agentic AI vulnerable?

A: The memory layer stores embeddings and contextual data that influence agent decisions. If compromised, attackers can subtly manipulate what the agent “believes,” leading to incorrect but seemingly legitimate actions.

Q: How can attackers exploit runtime workflows in agentic AI?

A: Attackers can influence planning inputs, prompts, or tool responses, causing agents to execute harmful actions using legitimate processes. These actions often appear normal in logs, making detection difficult.

Q: Why is observability a challenge in multi-agent systems?

A: Agentic systems operate at machine speed with continuous interactions between agents. Traditional logging systems are not designed to capture or process this level of dynamic, high-frequency activity.

Q: What are emergent failures in multi-agent environments?

A: Emergent failures occur when a small compromise in one agent or layer propagates across interconnected agents, resulting in large-scale issues that are difficult to trace back to the original source.

Q: What does effective recovery look like for agentic AI?

A: Effective recovery requires more than restoring data – it demands a coherent snapshot of all system layers, including memory, workflows, identities, and interactions, aligned to a verified trustworthy state.

Tim Zonca is Vice President, Portfolio Management, at Commvault.

More related posts


Thumbnail_Blog-What-is-Resops-2026

What Is ResOps – and Why Cyber Resilience Needs It

Read more about What Is ResOps – and Why Cyber Resilience Needs It
Thumbnail_Blog-Playbook-2026

Prove It Before You Need It: The Playbook I Wish I’d Had 10 Years Ago

Read more about Prove It Before You Need It: The Playbook I Wish I’d Had 10 Years Ago
Thumbnail_Blog-Medusa-is-Evolving-2026

Medusa Is Evolving. Cyber Resilience, Cyber Recovery, and ResOps Matter More Than Ever.

Read more about Medusa Is Evolving. Cyber Resilience, Cyber Recovery, and ResOps Matter More Than Ever.

Scaling a data-driven company is hard. Scaling one while meeting GDPR requirements, managing thousands of customers, enabling analytics teams, and standing up new infrastructure in under two weeks? That’s a different level of complexity.

In a recent episode of STRIVE, I sat down with Asif Dromi of monday.com and Ben Herzberg of Commvault to unpack what it really takes to operationalize data security at scale – not in theory, but in practice. This isn’t a high-level conversation about best practices. It’s a real-world look at how security, compliance, automation, and infrastructure decisions intersect when the clock is ticking.

Watch the full episode.

If you’re a CISO, data leader, architect, or compliance owner, this episode gives you something more valuable than theory. It shows how:

  • A fast-growing enterprise handled GDPR pressure without stalling innovation.
  • Infrastructure as code can simplify audits.
  • Automation reduces risk instead of increasing complexity.
  • Security and business agility don’t have to compete.

It’s rare to hear directly from operators who’ve done this under real constraints. That’s what makes this STRIVE conversation different.

Key Takeaways: Operationalizing Data Security at Scale

  • Compliance and growth don’t have to compete. Monday.com demonstrates how GDPR requirements and rapid expansion can coexist when security is built into architecture from the start.
  • Manual permissions don’t scale. Automation does. Infrastructure as code and API-driven access controls can turn governance from a bottleneck into a force multiplier.
  • Role-based access must evolve with data usage. As more teams depend on analytics, visibility and fine-grained controls become important to help prevent permission sprawl.
  • Operationalized security means visibility. It’s not just about setting policies – it’s about monitoring, auditing, and adapting controls dynamically as environments change.
  • Speed is possible when architecture is intentional. A compliant European data warehouse stood up in under two weeks because governance, automation, and tooling were designed to scale.
  • Security maturity enables innovation. When permissions, infrastructure, and compliance are programmable, organizations can move faster.

The Real Challenge: Growth + Compliance + Speed

For monday.com, the challenge wasn’t just storing European data in Europe. It was:

  • Enabling GDPR compliance and regional data residency.
  • Making sure employees only accessed relevant data.
  • Maintaining visibility and auditability.
  • Supporting analysts and developers who needed fast access.
  • Doing it all under intense business timelines.

As Asif explains in the episode, becoming a data-driven organization means internal access expands rapidly. The more teams rely on analytics, the more complex permissions become.

And that’s where many organizations hit a wall. Security becomes manual, permissions become fragile, and compliance becomes reactive. That’s not operationalized security. That’s a house of cards.

Designing Security into the Architecture from Day One 

One of the most compelling parts of the episode is how monday.com approached the problem architecturally. Instead of retrofitting compliance, it built:

  • A dedicated European data warehouse.
  • Clear role-based access controls.
  • Fine-grained permission models.
  • Automated governance layers.

Ben describes what happens in many large organizations: Over time, permissions accumulate in layers, often without central visibility. Eventually, no one is confident about who can access what. Operationalizing security means avoiding that drift. It means building systems where governance scales automatically as usage grows.

Automation Is the Force Multiplier

If there’s one theme that runs through this episode, it’s automation. Instead of treating permissions as tickets and manual updates, monday.com wrapped their infrastructure in code. Databases, roles, and access policies could be created and modified programmatically.

The result? A compliant, scalable environment stood up in less than two weeks. That’s not luck. That’s architecture. And it’s a powerful reminder that security doesn’t slow you down when it’s built correctly. It enables speed.

What Operationalizing Data Security Really Means

“Operationalizing” gets used a lot. In this episode, it’s defined as:

  • Continuous visibility into sensitive data.
  • Centralized and automated permission management.
  • Access tracking.
  • Integration with collaboration tools.
  • Policies that adapt as users and data grow.

Static controls don’t scale. Manual workflows don’t scale. Security must become dynamic – part of the operating fabric of the organization. And that shift is where many enterprises struggle today.

Watch the Full STRIVE Episode

In the discussion, you’ll hear more about:

  • How monday.com structured its European data warehouse.
  • The biggest lessons learned during rapid implementation.
  • Why automation was non-negotiable.
  • What companies often underestimate about permission sprawl.
  • How to think about operationalizing governance before AI initiatives expand.

Watch now.

FAQs 

Q: How can small teams implement scalable data security?

A: Start with a clear permissions model and infrastructure-as-code tools. Automate permission management early to help avoid manual bottlenecks as you grow.

Q: What role does automation play in compliance?

A: Automation helps enable consistency, reduce errors, and simplify audits. Using APIs and scripts, you can monitor and adjust permissions dynamically.

Q: How long does it typically take to set up a compliant, scalable data environment?

A: With the right planning and tools, organizations like monday.com have achieved this in less than two weeks. Speed depends on scope and existing infrastructure.

Q: What are best practices for operationalizing data security?

A: Implement role-based access controls, automate permission management, monitor access logs regularly, and integrate security tools with collaboration platforms for real-time oversight.

Chris Mierzwa is Senior Director, Portfolio Marketing, at Commvault.

More related posts


Thumbnail_Blog-GoogleWorkspace-2026

Expanding Google Workspace Protection with Commvault eDiscovery

Read more about Expanding Google Workspace Protection with Commvault eDiscovery
Thumbnail_Blog-Data-Leakage-Loops-2026

Are You Ready for Data Leakage Loops?

Read more about Are You Ready for Data Leakage Loops?
Thumbnail_Blog-Tornado-2025-Linkedin

The Trust Tightrope: Why New Yorkers Demand More from Businesses Than They Do from Themselves

Read more about The Trust Tightrope: Why New Yorkers Demand More from Businesses Than They Do from Themselves
Thumbnail_Blog_FinServ-Cybersecurity-2025

Modernizing Financial Cybersecurity: From Reactive to Resilient

Read more about Modernizing Financial Cybersecurity: From Reactive to Resilient

Key Takeaways

  • Compliance frameworks codify lessons learned from real-world failures and help organizations strengthen resilience, governance, and operational stability.
  • Organizations that approach compliance as a trust-building initiative can help strengthen customer confidence, partner relationships, and brand credibility.
  • Regulatory alignment and strong risk controls can help improve insurance outcomes by demonstrating a mature and resilient security posture.
  • Mapping compliance requirements to measurable business outcomes enables organizations to connect resilience investments directly to revenue protection and continuity.
  • Cyber resilience capabilities such as immutable backups, rapid recovery, and governance frameworks help organizations turn compliance into a competitive advantage.

In boardrooms across Europe and beyond, compliance has become a loaded word. It conjures images of endless documentation, mounting regulatory pressure, and the looming threat of fines.

GDPR. NIS2. DORA. The acronyms keep coming, and for many organizations, it can feel like they are choking on regulation.

But what if we’ve been looking at compliance the wrong way? What if compliance isn’t just about avoiding penalties – but about building a better, stronger, more resilient business?

The Insurance Analogy: Rules That Exist for a Reason

There’s a useful parallel between compliance and insurance.

When you insure your car, the insurer sets certain conditions. Your brakes must work. Your tires shouldn’t be bald. An alarm system might be required. You can argue about the inconvenience, or the cost – but fundamentally, those rules exist because they help reduce risk. They help make accidents less likely. They help protect both you and others.

And here’s the key point: Those requirements are usually a good idea, whether you buy the insurance or not.

Regulation works in much the same way. Governments and regulators don’t create frameworks because they enjoy it. Regulations are responses to real-world failures – data breaches, operational disruptions, systemic risk. They codify lessons learned the hard way.

You may object to the burden. You may find it frustrating. But when you look closely at what these frameworks require, it’s hard to argue that the core principles are unsound.

  • Protect customer data.
  • Enable operational resilience.
  • Know your supply chain risk.
  • Be able to recover from cyber incidents.
  • Demonstrate governance and accountability.

None of that is a bad idea.

From Avoiding Fines to Enabling Trust

Too often, compliance is framed defensively: “Do this so you don’t get fined.” “Do this so you don’t go to jail.”

That’s a low bar. And it’s a missed opportunity. When we shift the perspective, compliance becomes something much more powerful. It becomes a driver of trust.

Take GDPR as an example. At its heart, it’s about protecting personal data. If your organization implements strong data protection practices – not just to tick a box, but because your systems genuinely safeguard customer information – that builds trust. Customers are more confident doing business with you. Partners are more willing to integrate with you. Regulators view you as lower risk.

Trust is not a regulatory outcome. It’s a commercial advantage.

The same applies to the Digital Operational Resilience Act. It’s not just about reporting incidents; it’s about being able to withstand and recover from disruption. In a world where cyberattacks are inevitable, resilience is not optional. It’s foundational to continuity, reputation, and long-term value.

When compliance drives resilience, resilience drives business stability – and stability drives growth.

Regulation and Insurance: A Feedback Loop

There’s also a natural alignment between regulation and insurance markets. When regulators mandate certain standards, insurers quickly follow. Organizations that demonstrate compliance and strong risk controls are more attractive to underwriters. They may benefit from better terms, broader coverage, or more favorable premiums.

This creates a reinforcing cycle:

  • Regulation sets minimum standards.
  • Organizations strengthen their controls.
  • Insurers reward stronger risk postures.
  • Markets become more stable and resilient.

Compliance, in this context, becomes a signal to the market: We take risk seriously.

The Missing Link: Mapping Compliance to Business Outcomes

One of the most important opportunities for organizations – particularly technology providers – is to make the “line of sight” between compliance and business value explicit.

For example:

  • If a product creates immutable backups, that helps support regulatory requirements around data integrity.
  • If it enables rapid recovery from cyber incidents, that helps align with operational resilience mandates.
  • If it provides clear audit trails and reporting, that helps support governance and oversight requirements.

But it shouldn’t stop there. The next step is to articulate the business benefit:

  • Immutable backups help reduce the impact of ransomware – and protect revenue.
  • Faster recovery helps minimize downtime – and preserves customer confidence.
  • Strong governance helps reduce regulatory scrutiny – and enhances brand credibility.

This mapping is critical. Compliance is not the end goal; it’s the mechanism that enables the outcomes that businesses care about: continuity, reputation, customer trust, and competitive differentiation.

Compliance as Innovation, Not Obligation

There’s a tendency to treat compliance as a “get-it-done” exercise. A cost center. A necessary evil.

But if we look at history, many best practices that are now considered fundamental to modern IT and security originated in regulatory or insurance requirements. Over time, they became embedded in how well-run organizations operate.

Encryption. Access controls. Incident response planning. Business continuity testing. Third-party risk management.

At one time, these may have been viewed as regulatory burdens. Today, they are table stakes for any serious enterprise.

The organizations that treat compliance as an innovation catalyst – rather than a checkbox exercise – are often the ones that pull ahead. They embed resilience into their architecture. They design with governance in mind. They turn regulatory requirements into product capabilities and customer value propositions.

Cyber Resilience: Where Compliance and Strategy Converge

This is where cyber resilience becomes central.

Modern regulations increasingly recognize a simple truth: Prevention is not enough. Incidents will happen. The differentiator is how well an organization can respond and recover.

Cyber resilience – the ability to withstand, recover from, and adapt to cyber disruption – is no longer just a security concern. It’s a strategic imperative. It supports regulatory compliance, yes. But more importantly, it underpins operational continuity and business confidence.

When organizations invest in resilient architectures, immutable data, rapid recovery capabilities, and robust governance frameworks, they are not merely satisfying regulators. They are building durable enterprises.

A Different Conversation About Compliance

Perhaps it’s time to change the narrative.

Instead of asking, “What’s the minimum we need to do to comply?” we should be asking:

  • How does this regulation make us stronger?
  • What good practice is being codified here?
  • How can we use this to enhance trust with customers and partners?
  • Where does this create a competitive advantage?

Compliance done well is not about fear. It’s about foresight.

It reflects lessons learned across industries. It embeds best practice into everyday operations. And when connected clearly to product capabilities and business outcomes, it becomes a powerful commercial story.

Yes, regulation can feel burdensome. Yes, the acronyms keep coming. But underneath the paperwork lies something far more valuable: a framework for running a better business.

Compliance isn’t just about avoiding penalties. It’s about enabling resilience. And resilience, ultimately, is what drives sustainable success. Learn more about how Commvault enables data protection to help your organization meet compliance requirements here.

FAQs

Q: Why should organizations view compliance as more than a regulatory obligation?

A: Compliance frameworks often reflect best practices developed in response to real-world cyber incidents, operational failures, and governance challenges. Organizations that embrace compliance strategically can help strengthen resilience, improve trust, and create long-term business value.

Q: How does compliance contribute to customer trust?

A: Strong compliance practices demonstrate that an organization takes data protection, governance, and operational continuity seriously. This can help increase customer confidence, strengthen partner relationships, and position the organization as a lower-risk business.

Q: What is the connection between compliance and cyber resilience?

A: Modern regulations increasingly focus on an organization’s ability to recover from disruptions rather than solely preventing them. Investments in resilient infrastructure, immutable backups, and rapid recovery capabilities can help organizations maintain continuity during cyber incidents.

Q: How can compliance positively impact insurance and risk management?

A: Organizations with mature compliance programs and strong security controls are often viewed more favorably by insurers. This can lead to better coverage options, improved policy terms, and potentially lower premiums.

Q: Why is it important to connect compliance initiatives to business outcomes?

A: Compliance efforts are most effective when organizations clearly demonstrate how controls support broader goals such as protecting revenue, reducing downtime, and preserving customer trust. This helps leadership view compliance as a strategic investment rather than a cost center.

Q6: How can organizations turn compliance into a competitive advantage?

A: Businesses that embed resilience, governance, and security into their products and operations can differentiate themselves in the market. By proactively aligning with regulatory expectations, organizations can strengthen their reputation and create greater confidence among customers and stakeholders.

Darren Thomson is Field CTO at Commvault.

More related posts


Thumbnail_Blog-What-is-Resops-2026

What Is ResOps – and Why Cyber Resilience Needs It

Read more about What Is ResOps – and Why Cyber Resilience Needs It
Thumbnail_Blog-Playbook-2026

Prove It Before You Need It: The Playbook I Wish I’d Had 10 Years Ago

Read more about Prove It Before You Need It: The Playbook I Wish I’d Had 10 Years Ago
Thumbnail_Blog-Medusa-is-Evolving-2026

Medusa Is Evolving. Cyber Resilience, Cyber Recovery, and ResOps Matter More Than Ever.

Read more about Medusa Is Evolving. Cyber Resilience, Cyber Recovery, and ResOps Matter More Than Ever.

Key Takeaways

  • Operational sovereignty focuses on who can access systems and under which jurisdictions they operate.
  • Vendor access, telemetry flows, and support pathways can create hidden sovereignty gaps.
  • Operational sovereignty is harder to certify because it requires continuous visibility and auditing.
  • Organizations must be able to demonstrate and document every access pathway into sovereign environments.

Ask most organizations where their sovereignty program is strongest, and the answer is usually some version of the same two things: data locality and encryption. They know where their primary data lives. They’ve implemented bring-your-own-key or hold-your-own-key arrangements. They can point to certifications.

Ask them who accessed their sovereign environment in the last ninety days, from which countries, and under which legal jurisdictions – and the confidence tends to evaporate.

Operational sovereignty is the hardest pillar to audit, the most likely to be underestimated, and the most common place where a sovereignty posture that looks solid on paper breaks down in practice. The Digital Sovereignty Readiness Report names it as one of the four pillars – this post goes further.

The question most organizations can’t answer: ‘Who accessed your sovereign environment in the last 90 days, from which countries, and under which legal jurisdictions?’

What Operational Sovereignty Actually Means

Operational sovereignty is not about where data lives. It’s about who runs the environment – and who can reach it. It covers three things that most sovereignty programs treat as implementation details rather than first-class concerns:

  • Personnel access and jurisdiction. Every person who can access your sovereign environment – for support, maintenance, monitoring, or incident response – operates under a defined legal jurisdiction. If a support engineer in a country subject to a foreign data access law can reach your systems, the sovereignty of your infrastructure is only as strong as that engineer’s legal exposure.

Most organizations, when they audit this for the first time, find at least one support pathway that crosses a jurisdiction boundary they hadn’t mapped.

  • Third-party and vendor access. Your sovereignty boundary extends to every vendor, managed service provider, and software platform with access to your sovereign environment. ITSM platforms, monitoring tools, SIEM systems – if these sit outside your sovereignty boundary but have access to data or metadata within it, you have a gap that data locality controls cannot close.
  • Telemetry, billing, and control-plane traffic. Data sovereignty programs focus on primary data. Operational sovereignty requires mapping where everything else goes: the telemetry your infrastructure generates, the metadata your monitoring systems collect, the billing data your provider processes. These flows can cross jurisdiction boundaries even when primary data doesn’t – and they are rarely mapped.

Why This Pillar Is Harder To Certify – and Why That Matters

Data locality is relatively straightforward to document. You can point to a storage region, a data residency agreement, a third-party audit. Operational sovereignty doesn’t have the same paper trail. There is no certification that guarantees the jurisdictional status of every support engineer who might access your environment.

This is precisely what makes it both the hardest pillar to audit and the most important to get right. It also connects directly to the minimum viable sovereignty challenge: applying the right operational controls to the right workloads requires knowing what those controls are – and operational sovereignty is where that knowledge is most commonly absent.

The Supply Chain Dimension

NIS2, which extends cybersecurity obligations across energy, transport, healthcare, and digital infrastructure sectors, now requires organizations to assess the cybersecurity practices of their technology suppliers. For sovereignty programs, this has a direct implication: vendor sovereignty posture is no longer a procurement nicety. It is an auditable requirement.

That means asking new questions of every provider in your sovereignty boundary: Where are your support personnel located? Under which legal jurisdiction do they operate? What happens to the access they have to my environment if your company is acquired by a non-EU entity?

What Good Looks Like

An operationally sovereign environment has four characteristics that can be demonstrated, not just documented:

  • Every access pathway into the sovereign environment is mapped – not just primary access, but vendor access, support access, and monitoring system access.
  • The jurisdictional status of every person or system with that access is documented and audited on a defined cadence.
  • Telemetry, metadata, and control-plane traffic flows are inventoried and either contained within the sovereignty boundary or explicitly assessed and accepted as out-of-scope.
  • The organization can answer the ninety-day access question – precisely, with evidence.

One more thing: Operational sovereignty doesn’t end at access control. If recovery requires personnel who operate outside your sovereignty boundary, the posture fails at the moment of an incident. That’s the subject of the fourth post in this series.

The Digital Sovereignty Readiness Report includes a direct assessment question on operational sovereignty.

FAQs

Q: What is operational sovereignty?

A: Operational sovereignty addresses who manages and accesses an environment, including personnel, vendors, and support systems. It extends beyond where data is stored.

Q: Why is operational sovereignty commonly overlooked?

A: Many organizations focus primarily on data location and encryption. Access pathways, support personnel, and telemetry flows are often not fully audited.

Q: How do vendors impact sovereignty posture?

A: Vendors and managed service providers may have access to sensitive systems or metadata. Their legal jurisdictions and operational practices can affect overall sovereignty compliance.

Q: Why are telemetry and metadata important?

A: Even if primary data remains local, telemetry and metadata may cross jurisdictional boundaries. These flows can create compliance risks if left unmanaged.

Q: What does a strong operational sovereignty model include?

A: It includes mapped access pathways, documented jurisdictional controls, audited vendor access, and visibility into all telemetry and metadata flows.

Alex Zinin is VP/GM, Managed Service Providers, at Commvault.

More related posts


Thumbnail-Digital-Sovereignty-4

Sovereign Data You Can’t Recover Isn’t Actually Sovereign

Read more about Sovereign Data You Can’t Recover Isn’t Actually Sovereign
Thumbnail-Digital-Sovereignty-2

Minimum Viable Sovereignty: Why the Right Posture Isn’t the Same for Every Organization

Read more about Minimum Viable Sovereignty: Why the Right Posture Isn’t the Same for Every Organization
Thumbnail-Digital-Sovereignty-1

You Don’t Have a Sovereignty Strategy. You Have a Residency Policy.

Read more about You Don’t Have a Sovereignty Strategy. You Have a Residency Policy.

Key Takeaways

  • Sovereign architectures often prioritize audits and access controls over recovery readiness.
  • Recovery personnel, backup systems, and key custody models can create sovereignty gaps during incidents.
  • Consistent controls across primary and recovery environments are essential.
  • Sovereignty-ready resilience requires tested recovery procedures under realistic conditions.

Picture the moment. The attack has already happened. The incident response team is assembling. Someone must decide which systems come back first, in what order, using the correct recovery points.

And then someone realizes: The personnel with recovery system access are based in a different country. Worse, the recovery environment itself (hosted in a cloud region, a partner datacenter, or a secondary site) was never subject to the same sovereignty controls as the primary data.

The practice wasn’t subject to the same sovereignty controls as the primary data. The regulator is asking for status. The clock is running.

This is the scenario most sovereign architectures were not designed for – and the one the Digital Sovereignty Readiness Report calls out directly: most sovereign applications are designed for the audit, not the incident.

Most sovereign applications are designed for the audit, not the incident. The difference becomes visible at the worst possible moment.

The Recovery Blind Spot in Sovereign Architecture

Sovereignty programs are built around access control – who can reach the data, under what authority, through what pathway. That architecture is necessary. It is not sufficient. And it connects directly to the operational sovereignty gaps explored in the third post in this series: If the people who run your environment operate outside your sovereignty boundary, that problem doesn’t disappear during an incident. It becomes the problem.

What access control leaves unanswered is the harder question: What happens after an incident, when recovery is not just a technical operation but a legally constrained one?

A ransomware attack on a regulated European organization doesn’t simply create a recovery problem. It creates a recovery problem that must be solved within a jurisdiction, using personnel with appropriate authorizations, against recovery points that can be demonstrated to be clean and uncompromised.

The sovereign architecture designed to protect the data can make recovery harder if resilience wasn’t built into the original design.

The Specific Failure Modes

The ways sovereign recovery architectures fail are predictable – and common:

  • Recovery personnel outside the sovereignty boundary. The engineers who know the recovery systems may operate in a different jurisdiction. Under pressure, using them is the path of least resistance. It is also a sovereignty violation at the moment it is least convenient to have one.
  • Backup infrastructure without matching controls. Primary sovereign environments are carefully controlled. Backup infrastructure – particularly older or secondary environments – is frequently not subject to the same sovereignty requirements. If recovery points are stored or processed outside the boundary, compliant recovery is not available from compliant infrastructure.
  • Key custody under crisis conditions. Hold-your-own-key arrangements are designed for normal operations. Under crisis conditions – with primary systems compromised and time pressure acute – the key custody model that works in a routine maintenance window may become an obstacle to recovery. If this hasn’t been tested, it’s an assumption, not a control.
  • Cross-environment governance gaps. Organizations operating across multiple sovereign tiers – which is most of them – often have strong controls in primary environments and weaker controls in secondary environments that are also part of the recovery path. Consistency across the full estate is what auditors will look for. Gaps in secondary environments become visible exactly when consistency matters most.

Why Sovereignty Controls Can Complicate Recovery

The same controls that make a sovereign environment defensible to an auditor can make it harder to recover from. Data movement restrictions that prevent unauthorized exfiltration also constrain recovery orchestration. Key custody arrangements that ensure no provider can access your data without authorization also add friction when you need to restore quickly.

None of this means these controls are wrong. It means they have to be designed with recovery in mind from the start – not added to an architecture where recovery was an afterthought. This is the core of the minimum viable sovereignty principle: Calibrating controls to actual requirements includes recovery requirements, not just access control requirements.

What Sovereignty-Ready Resilience Requires

  • Clean recovery validation. Proving that recovery points are free from compromise before restoring to production – not just recent, but uncompromised. In a ransomware scenario, a recent backup may itself be compromised. The ability to identify and restore from a known-clean recovery point, validated before it’s needed, is a sovereignty requirement, not just a disaster recovery requirement.
  • Cross-environment governance. Consistent sovereignty controls and audit evidence across the full estate – not just the primary sovereign deployment. Every environment in the recovery path must meet the same requirements as the primary environment.
  • Tested under realistic conditions. Regular exercises that validate recovery under the conditions that will actually exist during an incident: the legal constraints that apply, the personnel who are available, the recovery points that are clean. An annual disaster recovery test that doesn’t account for sovereignty constraints is not a sovereignty-ready exercise.

The Question To Add To Your Sovereignty Review

There is a direct way to assess whether your recovery architecture meets the same sovereignty requirements as your primary data environment: Ask it as a question and require an honest answer.

Can you recover your sovereign data, cleanly, within defined tolerances, using personnel operating within your sovereignty boundary, right now – under real conditions, not a controlled exercise?

For most organizations, the honest answer reveals a gap. The organizations that find it now – before the incident – will be best prepared with evidence when the regulator asks for it. The ones that don’t will be building it under pressure, in front of the people they least want to disappoint.

The Digital Sovereignty Readiness Report includes a direct recovery architecture assessment question.

FAQs

Q: Why is recovery important to digital sovereignty?

A: Sovereignty is incomplete if organizations cannot recover data within the same legal and operational boundaries used to protect it.

Q: What are common sovereign recovery failures?

A: Common failures include recovery personnel operating outside the sovereignty boundary, backup infrastructure lacking matching controls, and inconsistent governance across environments.

Q: How can key custody complicate recovery?

A: Hold-your-own-key models strengthen security during normal operations, but they can slow recovery efforts during incidents if not properly tested.

Q: What is clean recovery validation?

A: Clean recovery validation confirms that recovery points are free from compromise before systems are restored. This is especially important in ransomware scenarios.

Q: How should organizations test sovereignty-ready resilience?

A: They should conduct realistic exercises that account for legal constraints, operational availability, and validated recovery points – not just standard disaster recovery testing.

Alex Zinin is VP/GM, Managed Service Providers, at Commvault.

More related posts


Thumbnail-Digital-Sovereignty-2

Minimum Viable Sovereignty: Why the Right Posture Isn’t the Same for Every Organization

Read more about Minimum Viable Sovereignty: Why the Right Posture Isn’t the Same for Every Organization
Thumbnail-Digital-Sovereignty-3

The Pillar Most Sovereignty Strategies Forget

Read more about The Pillar Most Sovereignty Strategies Forget
Thumbnail-Digital-Sovereignty-1

You Don’t Have a Sovereignty Strategy. You Have a Residency Policy.

Read more about You Don’t Have a Sovereignty Strategy. You Have a Residency Policy.

Key Takeaways

  • Minimum viable sovereignty (MVS) focuses on applying the right level of control to the right workloads.
  • Treating all workloads equally can lead to unnecessary complexity and costs or insufficient protection.
  • Organizations typically fall into three sovereignty profiles: true sovereign, regulated enterprise, and hybrid multi-cloud.
  • Consistent governance across mixed environments is one of the biggest operational challenges.

There is a version of the digital sovereignty conversation that leads organizations somewhere expensive, operationally burdensome, and – if they’re being honest – further than their actual obligations require. Maximum sovereignty sounds responsible. In practice, it’s often a miscalibration.

There is an equally common version that leads somewhere dangerously thin – controls that satisfy a checklist but wouldn’t survive an audit, an incident, or a regulator who has stopped accepting documented intent as proof of demonstrated control.

The organizations that get sovereignty right tend to do something more rigorous and more practical than either extreme: They ask what they actually owe, to whom, and for what. Then they build to that standard – no more, no less.

This is the discipline of MVS, introduced in the Digital Sovereignty Readiness Report  and developed in full here.

MVS isn’t a shortcut. It’s a recognition that the goal is the right level of control, applied consistently, across every workload that requires it.

Not All Workloads Are Equal

The starting point for an MVS approach is workload classification – and most organizations skip it entirely.

A trading system processing regulated financial data carries fundamentally different sovereignty obligations than an internal HR collaboration tool. A database holding personal data of EU citizens is subject to a different legal and regulatory regime than a development environment running anonymized test data.

Treating all of these identically – either by applying maximum sovereign controls across the board or by assuming a single deployment model covers everything – is how organizations end up either over-engineered or under-protected.

The right question before any deployment decision: What does this workload require across each of the four sovereignty pillars? The Readiness Report includes a self-assessment structured around exactly that question.

The Three Profiles – and What They Actually Need

Regulated enterprises fall into three recognizable profiles, each with different primary drivers and investment priorities.

  • The True Sovereign. Government agencies, defense contractors, and critical national infrastructure operators. For these organizations, sovereignty is not a compliance requirement – it is an operational mandate. Maximum control over every dimension of the technology stack is often legally required, and the cost tradeoffs are accepted because the alternative is not.
  • The Regulated Organization. Financial services firms, healthcare organizations, energy companies. These organizations face binding requirements from DORA, NIS2, GDPR, and sector-specific frameworks. Compliance obligations may also map to EU certification schemes – including EUCS, EUCC, BSI C5, and SecNumCloud – depending on sector and deployment context.

on-negotiable in certain areas – particularly around data residency, operational access controls, and recovery within jurisdictional boundaries. But not every workload carries the same obligation.

  • The Hybrid Multi-Cloud Organization. Organizations with existing hyperscaler investments facing increasing sovereignty pressure from customers, regulators, or procurement requirements. Their challenge is not wholesale migration – it’s layering sovereign controls onto a mixed estate and maintaining consistent governance across it.

The Cost of Getting Calibration Wrong

Over-engineering sovereignty creates its own operational risks. Organizations that apply maximum sovereign controls to workloads that don’t require them absorb cost and complexity that serves no regulatory or business purpose.

Under-engineering is the more common failure mode, and the more dangerous one. It typically doesn’t show up until the audit arrives – or, more seriously, until an incident occurs and recovery becomes a legally constrained problem. (That failure mode is the subject of the fourth post in this series.)

A Practical Starting Point

An MVS approach follows three steps:

  1. Classify workloads by their actual sovereignty requirements across each pillar – don’t start with deployment models.
  2. Map each workload class to the deployment tier that meets those requirements, across the full spectrum from public hyperscaler regions to sovereign public cloud to on-premises managed environments.
  3. Govern the resulting mixed estate consistently – controls, audit evidence, and recovery capabilities must be demonstrable across the full environment, not just the most-sovereign tier.

The third step is where most programs struggle. Maintaining consistent sovereignty controls across a mixed estate is an operational governance challenge – and specifically the domain of Operational Sovereignty – the subject of the third post in this series, the pillar most strategies treat as an afterthought.

Use the self-assessment in the Digital Sovereignty Readiness Report to locate your current posture across all four pillars.

FAQs

Q: What is minimum viable sovereignty (MVS)?

A: MVS is the practice of applying sovereignty controls based on actual business and regulatory needs. It is intended to help avoid both over-engineering and under-protection.

Q: Why is workload classification important?

A: Different workloads carry different regulatory and operational obligations. Classifying workloads helps organizations apply the appropriate level of sovereignty controls.

Q: What are the three common sovereignty profiles?

A: The three profiles are true sovereign organizations, regulated organizations, and hybrid multi-cloud organizations. Each has distinct operational and compliance requirements.

Q: What risks come from over-engineering sovereignty?

A: Excessive controls can increase operational complexity and costs without delivering meaningful compliance or business value.

Q: Why do mixed environments create governance challenges?

A: Organizations often operate across multiple cloud and infrastructure models. Maintaining consistent controls, audit evidence, and recovery standards across all environments is difficult.

Ruben Renders is Solutions Director, MSP, at Commvault.

More related posts


Thumbnail-Digital-Sovereignty-4

Sovereign Data You Can’t Recover Isn’t Actually Sovereign

Read more about Sovereign Data You Can’t Recover Isn’t Actually Sovereign
Thumbnail-Digital-Sovereignty-3

The Pillar Most Sovereignty Strategies Forget

Read more about The Pillar Most Sovereignty Strategies Forget
Thumbnail-Digital-Sovereignty-1

You Don’t Have a Sovereignty Strategy. You Have a Residency Policy.

Read more about You Don’t Have a Sovereignty Strategy. You Have a Residency Policy.

Key Takeaways

  • Data residency addresses where data is stored, but digital sovereignty also requires control over access, operations, and proper understanding of jurisdictional implications.
  • Operational sovereignty is often the weakest and least-audited part of most sovereignty programs.
  • A complete sovereignty posture depends on four pillars: data locality, technological sovereignty, operational sovereignty, and jurisdictional sovereignty.
  • Sovereignty is not binary; organizations must define a posture aligned to their regulatory and operational obligations.

Here is a question worth sitting with: When your organization made its sovereignty decision, what exactly did it decide?

For most, the answer is some version of the same thing. Pick a region. Move the workloads. Choose a cloud provider with data centers in-country. Check the box. The question of where data lives was answered, and the sovereignty conversation was considered closed.

But it wasn’t closed. It had barely started.

Data residency answers one question: Where? Digital sovereignty asks three more – who, how, and under what conditions?

The conflation of residency with sovereignty is understandable. Hyperscalers have made region selection feel like a sovereignty decision. Compliance checklists ask where data is stored. Regulatory guidance, at least in its earlier iterations, focused heavily on geography.

Choosing a sovereign cloud region is a real thing – it matters, it has operational implications, and it’s a necessary first step. But it is only a first step. And most organizations stopped there.

What Residency Doesn’t Answer

Think of it this way: Choosing a sovereign cloud region is like buying a safe. It tells you where your valuables are stored. It says nothing about who has a copy of the combination, who manufactured the safe, which country’s laws govern the manufacturer, or whether you can open it if compelled to.

Region selection answers one question. Three more remain entirely open – and these are the questions regulators, procurement committees, and auditors are now asking with increasing precision:

  • Who can operate your environment, and from where? Whether your cloud provider’s support personnel are subject to foreign jurisdiction is a sovereignty question that data residency cannot resolve. A routine maintenance window performed by a support engineer in a different legal jurisdiction is an access pathway your residency policy doesn’t cover. This is the domain of Operational Sovereignty – the hardest pillar to audit and the most commonly overlooked.
  • Under what legal regime can your data be accessed? A foreign technology provider operating infrastructure in-country does not automatically remove the reach of their home jurisdiction’s law. The extraterritorial reach of foreign legal regimes is a risk that geography alone cannot eliminate.
  • Can you recover your data if something goes wrong? Most sovereignty programs are built around access control. Very few address recovery – whether your data can be restored cleanly, within defined tolerances, by personnel who operate within your sovereignty boundary. That gap is where sovereignty postures most commonly fail under real conditions.

The Framework that Fills the Gap

A complete sovereignty posture spans four interdependent pillars. The Digital Sovereignty Readiness Report – available at readiverse.com – walks through each in full. In brief:

  • Data locality addresses where data and metadata actually travel.
  • Technological sovereignty covers control over encryption, key custody, and architecture portability.
  • Operational sovereignty covers who runs the environment and from where.
  • Jurisdictional sovereignty establishes the legal framework governing and affecting all of the above.

No single pillar is sufficient. A strong data locality posture with weak operational controls is not sovereignty – it is residency with unexamined risk.

What makes the framework useful is not its complexity. It’s the questions it generates. When an organization maps its current posture against all four pillars for the first time, it almost always finds gaps it didn’t know were there – not because the controls are absent, but because the questions were never asked.

Sovereignty Is a Sliding Scale

One more thing worth naming: Sovereignty is not a binary state. There is no certification that grants it and no single deployment model that guarantees it. It is a posture – a set of deliberate, auditable decisions. And the right level of that posture varies by organization, by workload, and by what you actually owe regulators and customers.

That calibration is what minimum viable sovereignty is about – the subject of the second post in this series.

Regulatory confidence is built long before the audit itself – through clearly defined requirements, not assumptions tied to geography.

Download the Digital Sovereignty Readiness Report for the four-pillar framework and a practical self-assessment tool.

FAQs

Q: What is the difference between data residency and digital sovereignty?

A: Data residency focuses on where data is physically stored. Digital sovereignty goes further by addressing who can access the data, how systems are operated, and exposure to which jurisdictions may create legal risk.

Q: Why is region selection not enough for sovereignty?

A: Choosing a cloud region only addresses geography. It does not resolve issues related to operational access, legal risks exposure, or recovery capabilities.

Q: What are the four pillars of digital sovereignty?

A: The four pillars are data locality, technological sovereignty, operational sovereignty, and jurisdictional sovereignty. Together, they create, what we believe, is a more complete framework for assessing sovereign readiness.

Q: Why is operational sovereignty difficult to manage?

A: Operational sovereignty involves monitoring who can access systems, where they operate from, and under which legal regime. These controls are harder to audit than simple data location requirements.

Q: Is digital sovereignty a fixed certification?

A: No. Sovereignty is an ongoing posture based on deliberate, auditable decisions that vary by organization, workload, and regulatory environment.

Ruben Renders is Solutions Director, MSP, at Commvault.

More related posts


Thumbnail-Digital-Sovereignty-3

The Pillar Most Sovereignty Strategies Forget

Read more about The Pillar Most Sovereignty Strategies Forget
Thumbnail-Digital-Sovereignty-4

Sovereign Data You Can’t Recover Isn’t Actually Sovereign

Read more about Sovereign Data You Can’t Recover Isn’t Actually Sovereign
Thumbnail-Digital-Sovereignty-2

Minimum Viable Sovereignty: Why the Right Posture Isn’t the Same for Every Organization

Read more about Minimum Viable Sovereignty: Why the Right Posture Isn’t the Same for Every Organization

Key Takeaways

  • Vishing attacks have surged dramatically, with organized groups industrializing social engineering to gain initial access through help desks.
  • Attackers quickly pivot from compromised human accounts to persistent machine identities like OAuth tokens and service accounts.
  • Most organizations lack governance and visibility over non-human identities (NHI), creating a major security blind spot.
  • Effective readiness depends on correlating identity signals and treating machine identities as high-risk assets.
  • True resilience requires the ability to detect and roll back unauthorized privilege changes before attackers establish persistence.

Your help desk staff just got a phone call. The caller knew the employee’s name, their manager, and the last four digits of their badge number. They asked for a password reset. Standard procedure. The IT rep complied.

That call was a fraud. And the attacker is now inside.

Voice phishing – vishing – jumped 449% in 2025. Adversary groups have turned social engineering into a scalable operation: recruiting callers, writing scripts, and paying $500 to $1,000 per successful help desk impersonation. They’re not looking for your data. They’re looking for a foothold.

Once inside, attackers don’t linger on the human account. They move laterally – stealing OAuth tokens, creating new administrative service accounts, embedding access in machine-layer credentials that nobody watches. Unlike human passwords, those credentials are rarely rotated. They don’t trigger login alerts. They can survive a full remediation of the original compromised user.

By the time your security team closes the ticket on the help desk incident, the attacker may have been quietly persistent in your environment for weeks. The governance gap makes it worse.

Fewer than 25% of organizations have formal policies for creating or decommissioning NHIs – the service accounts, API keys, and OAuth tokens that now outnumber human users by 144 to 1. Nearly all of them carry permissions far beyond what their function requires.

Most organizations have almost no confidence in their ability to detect an attack targeting this layer. That’s not a prevention failure. It’s a recovery planning failure.

What Readiness Looks Like

Prevention at the help desk matters – training, callback verification, out-of-band confirmation. But it isn’t enough on its own. Attackers are industrializing faster than awareness programs can keep pace.

Readiness means correlating the signals: A help desk interaction followed immediately by a multi-factor authentication (MFA) reset or a new token creation is a high-probability indicator of compromise.

It means treating machine identities as Tier 0 assets – governing their creation, scoping their permissions, and monitoring for unauthorized escalation. And it means having the ability to detect and roll back malicious privilege changes quickly, before they become the new normal.

Explore how Commvault identity resilience supports rapid detection, rollback, and recovery of your identity environment.

FAQs

Q: What is a vishing attack in the context of enterprise security?

A: Vishing (voice phishing) uses phone calls to impersonate employees and manipulate IT help desks into granting access – typically through password or MFA resets. It’s increasingly industrialized, with organized groups recruiting callers and using pre-written scripts to maximize success rates.

Q: Why do attackers pivot to machine identities after a vishing entry?

A: Human accounts get remediated. NHIs – OAuth tokens, service accounts, API keys – are more persistent and rarely rotated, often invisible to traditional monitoring. Migrating access to the machine layer allows attackers to maintain that persistence long after the original human credential breach is detected and closed.

Q: What does “identity resilience” mean in practice?

A: It means your organization can help detect unauthorized privilege changes in near real time and help restore the identity environment to a trusted state quickly. Detection alone isn’t sufficient – the ability to roll back malicious activity and verify that machine identities haven’t been tampered with (or if tampered with, to be rolled back to a prior good point in time) is what separates readiness from exposure.

Vidya Shankaran is Field CTO at Commvault.

More related posts


Thumbnail_Blog-Identity-Resilience-MachineID-2026-Linkedin

The Machine Identity Blind Spot Is Now a Primary Attack Surface

Read more about The Machine Identity Blind Spot Is Now a Primary Attack Surface
Thumbnail_Blog-Help-Desk-2026-Linkedin

When the Help Desk Becomes the Front Door to Your Entire Network

Read more about When the Help Desk Becomes the Front Door to Your Entire Network
Thumbnail_Blog-SHIFT-Identity-Resilience-2026-Linkedin

Your Identity Infrastructure Is a Target. Here’s What Commvault Is Doing About It.

Read more about Your Identity Infrastructure Is a Target. Here’s What Commvault Is Doing About It.
Thumbnail_Blog-Rise-of-AI-Agents-in-Resops-2026

Commvault and Microsoft: The Rise of AI Agents in ResOps

Read more about Commvault and Microsoft: The Rise of AI Agents in ResOps
Thumbnail_Blog_Resilient-Against-the-AI-Machine

Resilient Against the AI Machine

Read more about Resilient Against the AI Machine

Key Takeaways

  • Help desk social engineering is now a primary entry point, with vishing (voice phishing) attacks rapidly increasing and leading to credential compromise.
  • Non-human identities like service accounts and tokens are a major security blind spot, often unmanaged and heavily exploited for lateral movement.
  • Active Directory (AD) is a high-value target because of its centralized control and potential misconfigurations.
  • Prevention alone is insufficient; organizations need strong detection and rapid recovery capabilities to limit damage.
  • Immediate operational actions – like auditing accounts and correlating help desk activity with identity changes – can significantly reduce risk.

AD remains a primary target for attackers because it sits at the center of enterprise identity. Recent research shows that 67% of incidents now involve identity-related compromise, with attackers going after critical systems like AD within hours of initial access. Once compromised, recovery can take days or weeks – causing significant business disruption.

The question worth asking isn’t whether AD is a target. It’s how attackers get there – and why the path is so much shorter than security teams might expect.

3 Steps to Full Compromise

Adversary groups like ShinyHunters and Scattered Spider have turned social engineering into a production operation. Voice phishing – vishing – jumped 449% in 2025. Callers are recruited, scripted, and paid up to $1,000 depending on success and hit rate.

That means, it’s possible to start an attack with one step: Get a password reset or multi-factor authentication (MFA) change. That’s it.

From that single credential, the attacker moves laterally into cloud and virtualized environments. They harvest OAuth tokens, create new administrative service accounts, and embed access in machine-layer credentials. These non-human identities – service accounts, API keys, tokens – now outnumber human users 144 to 1. Sprawl and operational overhead makes rotation and audit difficult.

That lateral movement has a destination: Active Directory.

AD Is the Target

AD is the central nervous system of enterprise identity. Control it and you control everything – user accounts, group policies, and access to every domain-joined system in the network. The reason it’s so attractive to attackers – and so difficult to defend – is structural. Any authenticated user can read the entire directory. Every domain-joined system inherits trust from it.

Group Policy Objects linked at the domain head can be weaponized to disable security controls outright. Legacy protocols left enabled for application compatibility provide straightforward access. Microsoft’s own documentation says that “most identity attacks utilize common misconfigurations in Active Directory.”

When an attacker reaches the AD, they don’t need to force entry. The door is usually open.

Prevention Is Necessary but Not Sufficient

The standard security stack – MFA, endpoint detection, email filtering – is built around human behavior. It wasn’t designed to govern the machine identity layer or to detect the kind of slow, legitimate-looking privilege escalation that characterizes modern AD attacks. An attacker that moves from a compromised human account to a service account to a domain administrator over 72 hours may never trigger a single alert.

This is why the conversation must shift from prevention-first to recovery-first.

Prevention still matters. Least-privilege access, auditing AD changes, hardening default configurations, disabling inactive accounts – these can help reduce the attack surface. But given that half of organizations have already experienced an AD attack, designing only for prevention means designing to fail.

True identity resilience requires the ability to detect unauthorized privilege escalations in near real time, roll back malicious changes before they propagate, and restore the identity environment to a known-trusted state quickly – not in days or weeks, but fast enough to contain the blast radius. That means treating AD and the non-human identity layer as Tier 0 assets, with the same governance and recovery investment you’d apply to any other mission-critical system.

What To Do Right Now for Identity Resilience

The gap between where most organizations are and where they need to be on identity resilience is real. But it’s closeable. The immediate priorities are unglamorous and operational:

  1. Audit what’s in your AD.
  2. Find the accounts that shouldn’t still exist.
  3. Rotate the credentials that haven’t been touched in years.
  4. Correlate help desk activity against token- and account-creation events.

A help desk interaction followed by an MFA reset followed by a new service account is a high-confidence attack signal – and it’s detectable if you’re looking for it.

The longer-term work is architectural: Build recovery capability into your identity program so that when an attack succeeds – and it’s usually when, not if – you can contain it, reverse it, and try to restore trust faster than the attacker can consolidate their position.

Attackers are counting on your AD being ungoverned, your machine identities being invisible, and your recovery plan being theoretical. Close one of those gaps this quarter. Close all three and you’ve fundamentally changed the math. 

Learn how Commvault Cloud delivers comprehensive AD protection – from vulnerability assessment to one-click rollback and full forest recovery.

I recently joined Vidya Shankaran on the STRIVE podcast to talk about the governance gap for non-human identities. Check out our episode here. And be sure to read Vidya’s blog, The Machine Identity Blind Spot Is Now a Primary Attack Surface.

FAQs

Q: Why are help desks becoming a major security risk?

A: Help desks are often trusted to reset passwords and modify MFA settings, making them attractive targets for social engineering. Attackers exploit this trust to gain initial access with minimal resistance.

Q: What role do non-human identities play in attacks?

A: Sprawl and operational overhead make rotation and audit of non-human identities, such as service accounts and API keys, difficult. Attackers use them to maintain persistence and move undetected across systems.

Q: Why is AD such a critical target?

A: AD controls authentication and access across the network. Gaining control of it allows attackers to manage users, policies, and systems at scale.

Q: Isn’t MFA and endpoint security enough to stop these attacks?

A: These tools focus on human behavior and may not detect slow, legitimate-looking privilege escalation. Attackers can operate within normal patterns and avoid triggering alerts.

Q: What does a recovery-first security approach mean?

A: It means preparing for the reality that breaches will happen and prioritizing the ability to detect, contain, and reverse them quickly. This approach helps reduce downtime and can help limit overall impact.

Q: What are the most important steps to take immediately?

A: Start by auditing your AD, removing unnecessary accounts, rotating old credentials, and monitoring for suspicious sequences of help desk and identity-related activities.

Dan Conrad is Principal Technologist and Field CTO at Commvault.

More related posts


Thumbnail_Blog-Okta-Early-Access-2026

Commvault® Extends Identity Resilience to Okta

Read more about Commvault® Extends Identity Resilience to Okta
Thumbnail_Blog-Lateral-Access-2026

Staying Resilient Against Lateral Access Exploits

Read more about Staying Resilient Against Lateral Access Exploits
Thumbnail_3_AD_Blogs_2025

Active Directory Forest Recovery: Why Manual Methods Are No Longer Viable

Read more about Active Directory Forest Recovery: Why Manual Methods Are No Longer Viable
Thumbnail_6_AD_Blogs_2025

AD Recovery Testing: How to Know Your Recovery Plan Will Actually Work

Read more about AD Recovery Testing: How to Know Your Recovery Plan Will Actually Work

Key Takeaways

  • Non-human identities (NHIs) now vastly outnumber human users and are growing at a much faster rate, creating a significant and under-governed attack surface.
  • Attackers increasingly use social engineering, like voice phishing (vishing), to bypass human defenses and gain access to machine-layer credentials.
  • Most NHIs operate with excessive permissions and lack proper lifecycle management, contributing to accumulated “identity debt.”
  • Traditional security tools fail to detect threats in the machine layer because NHIs behave differently from human users.
  • Organizations must shift from prevention-first strategies to recovery-first approaches, prioritizing rapid detection and rollback of identity-based attacks.

For the past decade, enterprise security investment has followed the human. Better authentication. Stronger multi-factor authentication (MFA). Phishing simulation. Identity-centric architecture. These investments were the right response to the threat landscape at the time.

The threat landscape has moved.

Today’s most sophisticated adversaries aren’t trying to defeat your MFA. They’re using it as a door. A convincing phone call to your IT help desk, an MFA reset, and a compromised human account – that’s the entry. What they’re actually after is what’s behind it: the sprawling, under-governed layer of NHIs that connects every system in your environment.

The Scale of the Problem Is Staggering

Service accounts, API keys, OAuth tokens, AI agents – NHIs now outnumber human users by a ratio of 144 to 1, and they’re growing 4 to 10 times faster than human accounts. Yet fewer than 25% of organizations have formal policies governing their creation or decommissioning. Nearly all of them carry excessive permissions – rights that far exceed what their function requires.

This isn’t a new risk that suddenly appeared. It’s accumulated identity debt: years of provisioning without governance, automation without accountability, cloud expansion without visibility. And adversaries have noticed.

Vishing Is the Entry Point

Groups like ShinyHunters and Scattered Spider – operating under what researchers call the Scattered LAPSUS$ Hunters (SLH) cluster – have industrialized social engineering to exploit exactly this gap. Voice phishing rose 449% in 2025. These aren’t opportunistic calls. They’re coordinated operations: purpose-built scripts, recruited callers, financial incentives of up to $1,000 per successful help desk impersonation.

The call isn’t the attack. The call is the credential reset that gets an attacker past the human perimeter. The attack begins when they migrate to the machine layer – stealing OAuth tokens, creating administrative service accounts, embedding access into credentials that are rarely monitored and almost never rotated.

The human account gets remediated. The machine-layer access persists. The attacker has already moved on.

Three Vulnerabilities that Traditional Controls Can’t See

Standard security tools are designed around human behavior. They flag anomalous logins, unusual geolocation, suspicious email traffic. NHIs operate differently, and that difference is the blind spot.

OAuth abuse, for instance, looks like normal API traffic – even after a password reset. Thousands of undocumented service accounts operate in large enterprises with administrative privileges, often long after the projects that created them ended. Long-lived API keys embedded in DevOps pipelines carry broad access with no device context and no login alert.

MFA doesn’t cover them. Endpoint detection doesn’t see them. Email filtering is irrelevant to them.

The Framework Shift: From Prevention-First to Recovery-First

The logical response to a threat that often evades traditional detection is to stop assuming you can prevent every intrusion and start designing for rapid recovery from the ones that succeed.

That means treating NHIs as Tier 0 assets – with the same governance controls applied to domain administrators or cloud control planes managed with human identities. It means replacing static secrets with short-lived tokens and automatic rotation.

It also means correlating cross-domain signals: A help desk interaction followed by an MFA reset followed by a new token creation is a high-confidence indicator of compromise, and catching it early is the difference between containment and a prolonged breach. It means mapping NHIs to human identities for accountability.

Most importantly, it means having the capability to detect unauthorized privilege escalations and roll back malicious identity changes in real time – returning the environment to a known-trusted state before the damage extends.

Prevention still matters. But given the governance gap many organizations are carrying, recovery speed is becoming a primary resilience metric. Organizations should build identity programs designed for the attacks that are already happening, not the ones that were common five years ago.

Visit the Readiverse and check out our eBook The Non-Human Identity Crisis, which explores the full scope of the machine attack surface and the framework for identity resilience.

FAQs

Q1: What are non-human identities (NHIs)?

A: NHIs include service accounts, API keys, OAuth tokens, and AI agents that allow systems and applications to interact. Unlike human users, they often operate automatically and at scale, making them harder to monitor and control.

Q2: Why are NHIs considered a security risk?

A: NHIs often have excessive permissions and lack proper governance, making them attractive targets for attackers. Because they are rarely monitored or rotated, compromised credentials can persist undetected for long periods.

Q3: How do attackers exploit NHIs?

A: Attackers typically gain initial access through social engineering, such as voice phishing, then pivot to the machine layer. They steal tokens, create new service accounts, or embed persistent access in credentials that are not closely monitored.

Q4: Why don’t traditional security tools detect these threats?

A: Most security tools are designed to track human behavior, such as login anomalies or phishing attempts. NHIs generate normal-looking system traffic, which allows malicious activity to blend in with legitimate operations.

Q5: What is meant by a “recovery-first” security approach?

A: A recovery-first approach focuses on quickly detecting breaches and restoring systems to a trusted state rather than assuming all attacks can be prevented. This includes identifying unauthorized changes and rolling them back in real time.

Q6: How can organizations improve NHI security?

A: Organizations can treat NHIs as critical assets, implement strict governance policies, replace static credentials with short-lived tokens, and correlate signals across systems. Mapping NHIs to human owners also improves accountability and oversight.

Vidya Shankaran is Field CTO at Commvault.

More related posts


Thumbnail_Blog-SHIFT-Identity-Resilience-2026-Linkedin

Your Identity Infrastructure Is a Target. Here’s What Commvault Is Doing About It.

Read more about Your Identity Infrastructure Is a Target. Here’s What Commvault Is Doing About It.
Thumbnail_Blog-Rise-of-AI-Agents-in-Resops-2026

Commvault and Microsoft: The Rise of AI Agents in ResOps

Read more about Commvault and Microsoft: The Rise of AI Agents in ResOps
Thumbnail_Blog-Unified-Resilience-2026

Why AI Is Breaking Your Resilience Strategy (And What to Do About It)

Read more about Why AI Is Breaking Your Resilience Strategy (And What to Do About It)
Thumbnail_Blog_Resilient-Against-the-AI-Machine

Resilient Against the AI Machine

Read more about Resilient Against the AI Machine

Organizations today are building applications faster, automating workflows at scale, and turning data into insights, powered by platforms like Microsoft Power Platform. What started as a low-code productivity layer has quickly become mission-critical, embedded in the processes that support revenue generation, day-to-day operations, and strategic decision-making.

But as the reliance on these business intelligence assets grows, so does the associated risk. The same platform accelerating innovation can also amplify the impact of operational errors, misconfigurations, and malicious actions.

A misconfigured workflow, a deleted report, or a broken application can disrupt business processes, compromise decision-making, and erode trust in the systems the business relies on. And when something goes wrong, recovery is rarely straightforward.

Commvault is helping address these challenges with enterprise-grade data protection and recovery for Microsoft Power Platform, starting with Power BI – allowing organizations to help keep the insights, workflows, and apps they build protected and rapidly recoverable. 

Power BI: The Gap Between Insight and Recovery

At the center of many Power Platform deployments is Microsoft Power BI, providing analytics and business intelligence, transforming data into reporting, forecasting, and operational visibility.

When Power BI assets are lost or compromised teams can quickly lose access to trusted insights, interrupting reporting cycles, and delaying business decision-making.

In practice, however, protection strategies lag behind the importance of these assets. Many organizations rely on manual file exports or limited native capabilities that weren’t designed for comprehensive recovery. When something breaks, teams are often forced to rebuild manually with no ability to restore exactly what’s needed. This makes recovery slow, error-prone, and difficult to scale.

Commvault Cloud Backup & Recovery for Microsoft Power Platform

Now generally available, Commvault Cloud Backup & Recovery for Microsoft Power Platform helps organizations protect and recover their business-critical assets, such as reports, from accidental deletion, corruption, and malicious activity.

  • Automated, policy-based protection: Apply policy-driven backups across Power BI workspace assets, enabling consistent, scalable coverage without manual intervention.
  • Rapid, granular recovery: Restore individual reports, folders to a specific point in time, avoiding manual rebuilds and helping minimize downtime and disruption.
  • Isolated, immutable backups: Help protect data from ransomware and unauthorized changes with backups designed to prevent unauthorized modification or deletion.
  • Simplified compliance: Maintain long-term retention (up to 10 years), centralized audit logs, and reporting to support regulatory and internal requirements.

Unified Platform for Resilience

Commvault Cloud offers a unified platform to protect SaaS, cloud, and on-premises workloads, including Microsoft 365, Dynamics 365, Salesforce, VMs, databases, and endpoints. With Microsoft Power Platform support, customers can streamline protection, recovery, and resilience for more workloads, helping reduce tool sprawl and simplify operations.

How to Get Started

Commvault Cloud Backup & Recovery for Power Platform is delivered as a SaaS solution, designed for fast deployment and minimal operational overhead. Organizations can connect their Power BI environment, apply policy-based protection, and begin backing up critical data in a matter of steps.

Automated discovery protects new reports and folders are included as environments evolve, while centralized management provides a single place to monitor, manage, and recover data at scale.

What’s Next: Expanding Across Power Platform

We intend to expand protection and resilience across Microsoft Power Platform to include Power Apps and Power Automate, extending coverage to the applications and workflows that power your business. Plans, timelines, and features are subject to change and should not be relied upon in making purchasing decisions.

Protect What Powers Your Business

As reliance on Microsoft Power Platform grows, so does the need for resilient, enterprise-grade protection. With Commvault Cloud, you can:

  • Protect critical assets against deletion, corruption, and attack
  • Rapidly recover exactly what you need – without rebuilding everything
  • Maintain trust in data, decisions, and automation
Ready to make your Microsoft Power BI investment resilient?

Learn more and see Commvault Cloud in action at commvault.com/platform/power-platform.

More related posts


Thumbnail_Blog-What-is-Resops-2026

What Is ResOps – and Why Cyber Resilience Needs It

Read more about What Is ResOps – and Why Cyber Resilience Needs It
Thumbnail_Blog-Playbook-2026

Prove It Before You Need It: The Playbook I Wish I’d Had 10 Years Ago

Read more about Prove It Before You Need It: The Playbook I Wish I’d Had 10 Years Ago
Thumbnail_Blog-Medusa-is-Evolving-2026

Medusa Is Evolving. Cyber Resilience, Cyber Recovery, and ResOps Matter More Than Ever.

Read more about Medusa Is Evolving. Cyber Resilience, Cyber Recovery, and ResOps Matter More Than Ever.