Traditional identity and access management was built around a specific assumption: a human logs in, does some things, and logs out, at human speed, with a person accountable for each action. An AI agent breaks that assumption in almost every dimension, and the permissions gap this creates is one of the most consequential, least understood risks in production AI systems right now.

Why human IAM doesn't transfer cleanly

What proper agent IAM actually requires

Scoped, not inherited, permissions

Every agent should have access defined by what its specific task actually requires, not by whatever broader system or service account it happens to run under. This sounds obvious, yet it's consistently skipped under time pressure, because scoping permissions narrowly takes deliberate work and inheriting a broad existing credential doesn't.

Per-caller, per-context scoping

The same tool might reasonably need different access depending on which agent, in which context, is calling it. An agent running in an on-call engineer's session might warrant broader access than the same tool exposed to a general-purpose internal assistant used company-wide. This needs to be enforced at the point the tool is actually invoked, not assumed from a static role assignment.

Regular re-review, not one-time provisioning

Permissions granted at build time need a defined re-review cadence, tied to when the agent's capabilities or the systems it touches change, not a calendar date picked arbitrarily. An agent that gained new tool access six months ago and was never re-reviewed against its original permission scope is a common, quiet source of risk.

Full traceability

Every action an agent takes should be traceable back to the specific instruction, data source, and reasoning path that led to it. This is what makes a wrong action investigable after the fact instead of a mystery, and it's the foundation that both adversarial security testing and any future guardian-agent-style oversight depend on.

A related, harder problem
This overlaps directly with defending against prompt injection: an agent with narrowly scoped permissions has a much smaller blast radius even if it's successfully manipulated, since there's simply less it's authorized to do. Permission scoping and injection defense are two layers of the same underlying discipline, not separate problems.

Where this fits in the emerging "guardian agent" category

We've written about guardian agents, the emerging category of systems built specifically to provide continuous runtime oversight of AI agent identity and behavior. Proper permission scoping, traceable logging, and regular re-review aren't optional extras on top of that category; they're the foundation it depends on. You can build this foundation well before you ever consider a dedicated guardian-agent product.

How we approach this

We scope every tool an agent can access to the minimum required for its actual task, enforce that scoping at the point of invocation rather than through a static role, and build in the logging needed to trace any action back to its source. This is standard practice on every engagement with meaningful tool access, not an optional add-on.