Key Takeaways
- AI agent governance requires treating agents as a distinct identity class on the same platform that governs your machines and certificates, so you keep full visibility and control as AI scales across the enterprise.
- AI agents operate with autonomous decision-making, so identity governance becomes your primary control mechanism instead of behavioral prediction.
- Most identity tools only catch problems after they happen, but you need controls that govern identities from the moment they’re created.
- Multi-cloud environments and post-quantum cryptography are both reshaping the identity landscape, so a unified approach now prepares you for both.
Enterprise identity security has been shaped by a stable model for the better part of two decades. Users authenticate, systems authorize, and activity is logged and reviewed after the fact. That model drove the rise of original IAM platforms, and it worked because the underlying assumption held: identities were relatively stable, and their behavior was bounded enough to govern. That assumption is breaking.
The shift is not just that non-human identities now outnumber users. That has been true for years. What is changing is the nature of those identities. AI agents are not tied to a single workload or a fixed execution path. They initiate actions, request access dynamically, and interact across systems in ways that are difficult to predict ahead of time.
This is not just identity sprawl at scale. It is identity sprawl with autonomy.
In a recent Help Net Security interview, AppViewX CEO Archit Lohokare frames this as a structural shift, not an incremental one. The industry is not just dealing with more identities. It is dealing with a different class of identity altogether.
Why AI agents don’t fit your current identity system
Service accounts are predictable by design. They are created for defined workloads, operate within known boundaries, and can be constrained accordingly. AI agents operate with a different set of assumptions. They can request access at runtime, chain actions across multiple services, and adapt their behavior based on context. In many cases, they introduce variability that cannot be fully modeled ahead of time.
That variability is precisely why identity has to be the control plane. As AppViewX CEO Archit Lohokare explains, securing AI agents requires the same assume-breach mindset that redefined network security: assume the model is already hallucinating, and build governance from there. Knowing what an agent is, what it’s accessing, and when, and ensuring privileged access control governs what agents can actually do, is how organizations add determinism to an otherwise indeterminate system. Listen here as Lohokare goes deeper on this in his conversation with Primary Venture Partners.
| Attribute | Service Accounts | AI Agents |
|---|---|---|
| Creation | Pre-provisioned for a defined role | Instantiated dynamically as needed |
| Behavior | Predictable by design | Adaptive, context-driven |
| Access | Static, scoped at setup | Requested at runtime |
| Scope | Single workload | Chains actions across systems |
| Governance need | Periodic review | Continuous, identity-layer control |
That is not a semantic point. It forces a different approach to governance. Existing identity systems were built around two primary classes: humans and machines. Agents sit between the two; operating at machine scale, but with a level of decision-making variability that introduces new risk.
Treating them as an extension of either category creates blind spots. Treating them as a first-class identity category is the starting point for control.
Why identity becomes the control layer for AI
The deeper issue is not just that agents behave differently. It is that their behavior cannot be assumed to be correct, even when operating as designed.
One of the more useful reframes from the interview applies a familiar security principle to AI systems. Instead of “assume breach,” Lohokare suggests:
“You have to assume that the model… is already hallucinating.”
Taken seriously, that changes where control can be enforced. If the decision layer itself is unreliable, governance cannot depend on intent, expected behavior, or deterministic execution paths.
Control has to shift to something more consistent.
That is where identity plays a central role. It becomes the layer that defines what an agent can access, under what conditions, and with what level of privilege, regardless of how the request was generated.
Visibility gaps are no longer tolerable
Most enterprises are already operating with incomplete visibility into their machine identity landscape. Certificates, keys, and workload credentials are distributed across multiple systems, often with inconsistent ownership and fragmented governance.
That has been manageable if inefficient when identity behavior was relatively stable.
AI agents remove that margin for error.
They operate on top of existing credential infrastructure, often using identities that the environment already trusts. Without a unified, real-time inventory, organizations cannot reliably answer basic questions:
- What identities exist?
- What do they have access to?
- How is that access being used?
More importantly, they cannot enforce consistent policy because the underlying data is incomplete.
This is why lifecycle management becomes central rather than adjacent. Discovery, issuance, rotation, and revocation are no longer just operational tasks. They are foundational to maintaining control in an environment where identities are constantly created and used at speed.
At the same time, cryptographic change is becoming unavoidable. As organizations prepare for post-quantum transitions, the lack of visibility and lifecycle consistency will become an even bigger constraint.
Fragmented identity tools break down at scale
The industry response so far has been incremental. Endpoint platforms add identity context. Detection systems incorporate certificate visibility. Point solutions emerge for specific aspects of agent behavior.
Each improves visibility in isolation. None establish consistent control.
What organizations end up with is a fragmented identity landscape:
- Policy defined in multiple places
- Enforcement that varies by system
- Incident response that depends on stitching together disconnected signals
This approach does not scale in a world of autonomous agents and evolving cryptographic requirements.
The alternative is a more unified approach where identities across certificates, keys, workloads, and agents are governed with consistent policy and lifecycle controls.
Evaluating identity platforms
As the space evolves, feature sets are beginning to look similar across platforms. Many solutions can provide some level of visibility into non-human identities.
The difference shows up in how control is actually applied.
Solutions that focus primarily on observation provide useful signals but operate after identities are already in use. They can highlight risk, but have limited influence over how identities are created, managed, or retired.
In contrast, approaches that emphasize lifecycle governance can enforce policy earlier and more consistently across discovery, issuance, rotation, and revocation.
This becomes increasingly important as:
- Machine identities continue to grow in volume
- AI agents introduce more dynamic and unpredictable behavior
- Cryptographic standards evolve
Bringing AI agents into this model ensures they are governed as part of the broader identity ecosystem, rather than treated as a separate problem.
The next constraint is not AI adoption, it is control
AI adoption is accelerating across every part of the enterprise. The constraint is no longer whether organizations can deploy these systems. It is whether they can maintain control over them.
Lohokare points to what is coming next:
“The two biggest tectonic shifts… are artificial intelligence and post-quantum.”
Both trends converge at the identity layer.
- AI increases the number and variability of identities
- Post-quantum changes the cryptographic foundations those identities rely on
Together, they put pressure on systems that were not designed for this level of scale or change.
Together, they put pressure on systems that were not designed for this level of scale or change.
For CISOs, the implication is straightforward. The challenge is not enabling AI. It is ensuring that the identities behind these systems can be consistently discovered, governed, and controlled.
Organizations that take a more unified approach to identity now will be better positioned to scale, without losing visibility or accountability.
Those that do not will find themselves managing increasing complexity with fragmented tools and limited control.
Tags
About the Author
