A practical identity framework for AI agents takes shape

AI agents are now acting inside enterprise systems with delegated authority, and the identity tooling built for human staff does not cleanly cover them. That is the starting point of a practical framework published by The Hacker News on 28 September 2026, which sets out what identity and access management, or IAM, should look like when the actor holding the credential is software rather than a person.

The guide defines IAM for AI agents as the identity-control architecture that governs agents as first-class actors. According to the piece, agents authenticate, invoke tools, and act across enterprise systems on behalf of someone else. That combination is what makes them different from both human users and traditional service accounts: they hold delegated authority, and they use it dynamically as they call tools and move between systems.

The guide covers four areas. The first is the limits of conventional provisioning, the routine process of creating accounts and assigning permissions. The implication of the framing is that the usual joiner-mover-leaver model, designed around employees and their job roles, does not map neatly onto agents that may be created, used and discarded at machine speed, and whose permissions derive from a delegating principal rather than from a job title.

The second area is the components that matter in an agent identity architecture. The third is how to evaluate framework choices, in other words how a security team should compare the competing approaches now appearing on the market rather than adopting the first one offered.

The fourth is the one most likely to interest anyone who has to answer to an auditor: what runtime evidence proves an agent behaved as intended. This is a shift in emphasis. Conventional IAM largely answers the question of who was granted access. The guide's framing is that for agents, grant-time controls are not enough on their own, and organisations need evidence gathered while the agent is running that shows what it actually did with the authority it was given.

The publication is a guide rather than a standard or a regulatory requirement, and it does not carry the weight of a formal specification. But it reflects a problem that security teams are increasingly raising: agents are being deployed into production with credentials and tool access, while the controls governing them are still borrowed from a model built for human employees. The practical question it puts to enterprises is whether they can currently demonstrate, after the fact, that an agent stayed within the authority it was delegated.