AI agents are moving from pilots into workflows that can read data, make decisions and execute actions. That turns the security discussion from model safety alone to identity safety. An agent is a non-human identity with delegated authority, and its ability to act must be tied to a clear identity, purpose and policy.

Traditional access management assumes a human initiates a request and can be challenged or held accountable. Autonomous agents can operate continuously, chain tools and pass data between services. If an agent inherits a broad user session or uses a long-lived secret, compromise can become a privilege escalation path.

Establishing accountable machine identities

Every agent or agent workload should have a distinct machine identity rather than a shared service account. It should carry metadata describing its owner, function, environment, workflow version and approved tools.

Applying context-aware least privilege

Authorization must be granular. An agent that summarizes documents should not automatically delete records, change policies or create credentials. Policy should evaluate the requested action, target resource, context and risk, with approval for sensitive operations.

Monitoring identity behavior

Logs must show which agent acted, who or what delegated the action, which tools were called and what data crossed the boundary. This creates an auditable chain for Agentic Identity and makes identities visible, constrained and revocable.

Source: Computer Weekly