AI agent identity and access management means giving each agent its own identity, only the permissions its task requires, short-lived credentials, human approval for high-impact actions, and logs that tie every action to a named owner. An agent with tool access can read data, change records and call other systems without a person in the loop. Treat it as a privileged identity, not as a feature of the application that hosts it.
Why an agent is a privileged identity
Most identity programs were built for two kinds of principals: people, who log in and act deliberately, and services, which run fixed code. An agent is neither. It chooses which tools to call at run time, based on model output that can be wrong or manipulated.
OWASP's Top 10 for LLM Applications 2025 names this problem LLM06, Excessive Agency. It traces the risk to three root causes: excessive functionality, excessive permissions and excessive autonomy. It lists the triggers as hallucination, direct or indirect prompt injection, a compromised extension, and, in multi-agent systems, a malicious peer agent.
MITRE ATLAS, the adversary knowledge base for AI systems, shows why this matters for access control. Its technique AI Agent Tool Invocation (AML.T0053) notes that an agent may have tools a user cannot reach directly, so abusing the agent gives an adversary tools they would not otherwise have. In access control terms, the agent is a confused deputy with entitlements of its own.
That sets the scope here. Defenses against prompt injection and sandboxing reduce the chance of manipulation. Identity and access controls limit what a manipulated or mistaken agent can do. You need both, and the second works even when the first fails.
Give every agent its own identity
The first failure is the shared identity. OWASP's examples include an extension that only needs to read data but connects with an identity that also holds update, insert and delete rights, and a per-user extension that runs under a generic high-privileged identity. Both make it impossible to say which agent did what, and both hand the agent far more than it needs.
Each agent, and ideally each deployment of an agent, should have a distinct identity registered in your identity provider. That identity needs a named human owner, a stated purpose, a review date, and an environment boundary so a test agent never holds production rights.
NIST's National Cybersecurity Center of Excellence (NCCoE) has opened this question for public comment. Its draft concept paper, Accelerating the Adoption of Software and AI Agent Identity and Authorization, published in February 2026, asks what identification metadata an agent needs, whether agent identities should be ephemeral, and how keys are issued, updated and revoked. It names existing standards as building blocks, including OAuth 2.0 and 2.1, OpenID Connect, SCIM and SPIFFE/SPIRE.
It is a concept paper, not a standard, and its status after the comment period closed in April 2026 is unclear. Check before you build on it. The practical reading is that you do not need to wait for a new standard to stop sharing service accounts between agents. Workload identity systems already issue each software workload its own verifiable identity, and agents can use them today.
Apply least privilege to tools, not just to the agent
Least privilege for an agent has two layers. The first is which tools the agent can call at all. The second is what each tool can do once called.
OWASP's mitigations for Excessive Agency map to both layers. Minimize the extensions an agent can reach. Minimize the functionality inside each extension, so a tool that only needs to read a mailbox cannot also send mail. Avoid open-ended extensions such as a shell command or an arbitrary URL fetch, which let the model decide the action instead of choosing from a fixed menu. Minimize the permissions each extension holds in downstream systems.
Consider a hypothetical support agent that summarizes tickets and can also issue refunds. If both capabilities share one credential, a poisoned ticket that persuades the agent to refund an account uses a permission the summarizing task never needed. Split them into two tools with two scopes, and give the summarizing task only the first.
OWASP also recommends executing extensions in the user's context where an agent acts for a person, using OAuth with the minimum scope. That keeps the agent's reach no wider than the requester's own. Where an agent acts for no one in particular, give it a narrow service identity and record that choice in your register.
The UK NCSC's May 2026 post, Thinking carefully before adopting agentic AI, states the principle in plain terms: apply least privilege, and give agents only the minimum access they need, for the shortest time required.
Use short-lived, scoped credentials
Static keys are the weakest part of most agent deployments. MITRE ATLAS lists Credentials from AI Agent Configuration (AML.T0083), which covers API keys, tokens and connection strings stored in agent configuration files. An attacker who can read the configuration, or persuade the agent to reveal it, walks away with a credential that works from anywhere.
The NCSC post advises avoiding long-lived credentials, using temporary credentials where possible, and revoking higher-level access when a task is complete. In practice that means:
- Issue a token per task or per session, with an expiry measured in minutes or hours rather than months.
- Scope each token to one tool and one set of resources, not to a whole API.
- Keep secrets out of prompts, configuration files and tool output, and fetch them from a secrets manager at call time.
- Revoke on completion and on any anomaly, without waiting for expiry.
This follows the logic of NIST SP 800-207, Zero Trust Architecture, published in August 2020. It states that there is no implicit trust based on network location or ownership, and that authentication and authorization are discrete functions performed before a session to a resource is established. An agent running inside your network earns no trust from that position. Every call to every resource is authorized on its own.
The NCCoE paper raises the harder version of the problem: how to apply least privilege when an agent's actions are unpredictable, and how to update policy dynamically. A workable answer today is to enforce policy at the resource, not in the prompt. OWASP calls this complete mediation: authorize requests in the downstream system rather than trusting the model to decide whether a request is allowed. The model can be talked out of a rule written in a prompt. It cannot be talked out of a denied API call.
Require approval for high-impact actions
Not every action needs a human. Reading a public document does not. Moving money, deleting records, changing access rights, sending external communications and touching production systems do.
OWASP recommends requiring user approval for high-impact actions. Define "high-impact" before deployment, in a table that lists each tool, the action, whether it can be reversed, and whether approval is required. Reversibility is the most useful test: if you cannot undo the action, a person should approve it.
Two design points make approval real instead of ceremonial. First, the approver sees the actual action and its parameters, not a model-written summary of them. Second, the approval is bound to the identity of the person who gave it. The NCCoE concept paper asks how to bind an agent's identity to a human identity so that human-in-the-loop decisions are attributable.
The NCSC post adds an accountability point. Humans remain accountable for the decision to deploy an agent, the access it is granted, the safeguards around it and the consequences. Before connecting an agent to real systems, it says, define who owns it, who approves its access, who monitors it, who reviews incidents and who can stop it. Write those five roles down for every agent in your register.
Build an audit trail that survives an investigation
If you cannot reconstruct what an agent did and why, you cannot investigate an incident or satisfy an auditor. The NCSC's Guidelines for secure AI system development, version 1.0, published in November 2023 with international partners, tell operators to monitor and log system inputs to enable compliance obligations, audit, investigation and remediation. They also say to treat changes to data, models or prompts like new versions.
For agents, a useful log entry records:
- The agent identity and its human owner.
- The request that started the run, and the user on whose behalf it ran, if any.
- Each tool call, with parameters, the credential used and the result.
- Any approval requested, who granted it, and when.
- The version of the model, prompt and tool configuration in effect.
The NCCoE paper goes further and asks about tamper-proof, verifiable logging of actions and intent, and about non-repudiation. Those questions are open. Until they are settled, write logs to a store the agent cannot modify. OWASP also lists two damage limiters: logging and monitoring extension activity, and rate limiting. A cap on tool calls limits how much harm a runaway agent can do before someone notices.
The NCSC states the bar bluntly: if you cannot understand, monitor or contain an agent's actions, it is not ready for deployment.
Where to start
The joint guidance Careful adoption of agentic AI services, issued in spring 2026 by CISA, the NSA, Australia's ACSC, the Canadian Cyber Centre and the New Zealand and UK NCSCs, calls for a layered defense and strict access controls. The NCSC summary of it recommends tightly bounded pilots. These steps fit that advice:
- Inventory every agent that holds credentials, including those inside SaaS products and developer tools. Record its owner, tools, credentials and the systems it touches.
- Replace shared and long-lived credentials with per-agent identities and short-lived, scoped tokens. Start with the agents that can write or delete.
- List each tool and mark the actions that are irreversible or external. Put human approval in front of those actions, with the approver seeing the real parameters.
- Move authorization into the downstream systems so a denied call stays denied whatever the model outputs.
- Turn on logging that captures the five items above, send it to storage the agent cannot edit, and test that you can answer "what did this agent do last Tuesday?" within an hour.
- Pilot with one bounded task, review the logs after two weeks, and widen access only where the evidence supports it.
Work in this order: inventory first, then permissions, then approvals, then logging. Test each control by trying to make an agent exceed its scope.
References
- OWASP GenAI Security Project: LLM06:2025 Excessive Agency
- NIST: SP 800-207, Zero Trust Architecture
- NCSC (UK): Guidelines for secure AI system development, secure operation and maintenance
- NCSC (UK): Thinking carefully before adopting agentic AI
- NIST NCCoE: Accelerating the Adoption of Software and AI Agent Identity and Authorization (draft concept paper)
- MITRE ATLAS: ATLAS data, techniques AML.T0053 and AML.T0083
- Canadian Centre for Cyber Security: Joint guidance on careful adoption of agentic AI services





