AI agent security: where identity fits

Updated

AI security covers a lot of ground: model safety, training data, prompt injection, output filtering. Once an AI system can act — call tools, change records, move data — one part of it becomes the one that decides how bad a mistake can get: identity and access. An agent can only do damage with permissions it holds. AI agent security is largely the discipline of making those permissions narrow, visible, enforced and accountable.

The risks that come with agents

The OWASP Top 10 for Large Language Model Applications names the patterns security teams now see in practice. Two matter most once agents have tools:

  • Prompt injection. Text the agent reads — an email, a web page, a document — can carry instructions. A successful injection does not need to break the model; it only needs the agent to call a tool it already has permission to call.
  • Excessive agency. Agents are given more tools, broader permissions or more autonomy than their task needs. When something goes wrong, every extra permission becomes part of the blast radius.

Around those sit the identity risks every organisation already knows from service accounts, now at the speed of software: agents with no accountable owner, credentials that never rotate, access nobody reviews, and actions no one can reconstruct afterwards.

The identity controls that address them

You cannot prompt-engineer your way out of excessive agency. These controls work regardless of what the model is told:

  1. Treat every agent as an identity. Registered, typed, with an owner and a status — see what is a non-human identity?
  2. Least privilege per tool. Grant specific tools on specific servers, with short-lived scoped credentials, ideally bound to the agent’s key so a stolen token is useless elsewhere.
  3. Enforce at runtime. Put a checkpoint on every tool call that checks identity, scope, rate and data before the call lands — see what is an MCP gateway? Even a successfully injected agent can then only do what its grant allows.
  4. Review agent access. Agents belong in access reviews, with the context reviewers need: last use, how the access compares with similar agents.
  5. Separate duties. The same rules that stop one person both creating and approving a payment should stop one agent — or one agent and the person it acts for — from holding both halves of a sensitive process.
  6. Keep one tamper-evident audit trail for people and agents, so an incident review can establish exactly what happened.

The full set of practices is in how to govern AI agents.

How AtmanID approaches it

AtmanID is identity and access management built for this: humans and AI agents on one identity model and one tamper-evident audit trail (what is AtmanID, audit chain). Agents receive scoped tokens for specific MCP servers and tools, and a tenant can require them to be DPoP-bound to the agent’s key (MCP bearer-token flow). Every agent → MCP call passes through the AtmanID gateway for scope, rate-limit and DLP checks, starting in shadow mode (MCP gateway). Access is certified in campaigns with reviewer context (certification campaigns), and separation-of-duties rules block or warn on conflicting access (SoD rules).

Further reading: OWASP Top 10 for LLM Applications.