AI agents are exposing a gap in how enterprises define identity. Many organisations have invested heavily in human IAM, yet the most active identities in their environments may now be applications, workloads, connectors, bots, and autonomous agents. The security question is shifting from who is the user to which non-human identity is acting, under whose authority, and with what constraints.
Why existing controls fall short
Human-centric controls often assume predictable behaviour. A user authenticates, selects an application, and performs a limited number of visible actions. An agent can interpret an objective, choose a sequence of tools, and continue until the objective is complete. This makes the agent’s effective privilege larger than the permissions shown in a simple role assignment.
Risk also accumulates across connections. An agent may use one identity to read a ticket, another to query a database, and a third to update a cloud resource. Without correlation, each system sees a legitimate request while the organisation cannot see the full chain.
Identity security must follow the action
NHI security requires an identity that persists across the agent’s workflow. Every request should carry enough context to identify the agent, its sponsor, the task, and the execution environment. Policies can then evaluate whether the action is expected and whether the agent is entitled to use the requested tool.
Short-lived credentials reduce the impact of theft. Scoped tokens reduce the number of resources an agent can reach. Approval gates can be reserved for high-impact operations instead of slowing every low-risk action. These controls are more effective when they are issued dynamically rather than embedded permanently in configuration files.
Building practical assurance
Start by mapping the non-human identity estate. Include service accounts, API keys, workload identities, OAuth applications, and agents operated by business teams. Assign owners and record the purpose, data access, and dependencies of each identity.
Then measure behaviour. Look for unexpected tools, unusual volumes, new destinations, privilege escalation, and actions outside approved hours or workflows. Behavioural evidence should complement, not replace, strong identity proof.
The result is a more realistic IAM model. AI agents can be useful at scale, but only when their identities are visible, their permissions are bounded, and their decisions leave an accountable record.