COMMENTARY: For years, identity security revolved around people. When an employee joined an organization, they received accounts, permissions, and access to the systems they needed. When they left, IT disabled those accounts, revoked credentials, and collected devices.
AI agents challenge that model. They can be created in minutes, complete a task, and disappear just as quickly — or continue operating long after the employee who created them has left the organization.
[
SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]
As organizations increasingly adopt agentic AI, security teams need to think beyond managing human identities and consider how to manage AI agent identities throughout their lifecycle.
Accountability starts with ownership
One of the biggest mistakes organizations can make is treating AI agents as independent actors. Unlike employees, AI agents can't be held accountable for their decisions, so every AI agent should have a clearly defined human owner who remains accountable for what the agent is allowed to do and the actions it takes.
That human ownership doesn't replace the need for the agent to have its own identity. In fact, both are necessary: the agent needs a unique identity so its activity can be tracked, and a human owner so accountability never becomes disconnected from the organization's governance.
Related reading:
That sounds straightforward, but many organizations haven't established clear identities for AI agents. A recent Cloud Security Alliance studyagents dedicated identities. As a result, more than two-thirds say they can’t clearly distinguish between human and AI agent activity.
If organizations can't reliably identify who — or what — performed an action, accountability quickly becomes blurred.
If multiple agents inherit the same credentials, investigating suspicious activity becomes significantly harder. When an incident occurs, security teams need to determine not only what happened, but which agent performed the action and who authorized it.
AI may perform the work, but investigators should always be able to trace that work back to a human owner.
Ephemeral doesn’t mean untraceable
Maintaining that accountability becomes much more difficult when the AI identity itself no longer exists. That’s becoming a reality as organizations experiment with ephemeral agents designed to perform a single task before terminating themselves.
Traditional investigations assume the identity still exists. An employee can be interviewed, a service account still exists in Active Directory, and a server has logs.
Ephemeral AI agents challenge those assumptions. An agent may complete its task, terminate itself, and disappear long before anyone realizes something went wrong. If investigators discover suspicious activity hours or days later, they can't rely on examining the agent itself. The investigation has to rely on the evidence the organization preserved while the agent was alive.
That changes what organizations need to preserve. Logging should capture more than individual actions. Security teams need enough context to reconstruct an agent's entire lifecycle, including who created it, what permissions it was granted or inherited, which systems it accessed, what actions it performed, whether it spawned additional agents, and when it was terminated.
Correlating those events allows investigators to stitch together what happened, even after the agent itself no longer exists.
Offboarding is becoming more complicated
Employee offboarding has always been a security event because organizations need to ensure that former employees no longer have access to corporate systems or data.
Now organizations need to ask another set of questions: Did this employee create any AI agents? Are those agents still running? Should they continue running? Does anyone know they exist?
Those questions don’t always have obvious answers. Automatically shutting down every agent associated with an employee who is leaving may sound like the safest approach, until one of those agents turns out to support a critical business process. On the other hand, leaving unknown agents running indefinitely creates a different set of risks, since an undocumented agent with access to sensitive systems can easily become a hidden attack surface.
Without visibility into what agents exist, who created them, and what they're responsible for, organizations can't confidently decide whether an agent should be retired, reassigned, or allowed to continue operating.
Governance has to keep pace with autonomy
As organizations give AI agents greater autonomy, governance becomes increasingly important.
That starts with visibility. Security teams need to know where AI agents are running, what they're doing, and what data they're accessing before they can effectively govern them.
Organizations don't need entirely new governance principles for AI agents. They can extend the disciplines they've already developed for human and machine identities.
The identities may differ, but effective governance still depends on answering the same questions: Who owns it? What can it access? How is it monitored? And what happens when it's no longer needed?
Organizations that successfully apply their existing identity practices to AI agents will be better prepared to embrace greater AI autonomy without sacrificing accountability or control.