Access Control Patterns for Enterprise Agents
Give each agent its own identity, scoped short-lived tokens, per-action policy checks and an audit trail, so a wrong step stays small and traceable.
Answer first
Treat an AI agent as a new kind of user that can be tricked. Give it its own identity, let it act on a person's behalf only through scoped and short-lived tokens, check policy on every action rather than only at login, require approval for actions that are hard to undo, and log every call with the identity of the agent and the person behind it. The aim is not to prevent every mistake. It is to keep each mistake small and traceable.
Why agents need different controls
A human user with access to a CRM reads a record, decides, then edits it. An agent with the same access can read thousands of records in a minute and act on instructions hidden in any of them. The OWASP Top 10 for LLM Applications (2025) lists both problems: Prompt Injection (LLM01) and Excessive Agency (LLM06), where a system is given more functionality, permissions or autonomy than its task needs.
Classic access control assumes the caller's intent is stable for the session. With agents, intent can change mid-task because of content the agent read. That pushes checks from session start to each action.
Patterns that work
Give the agent its own identity
Register each agent (or each agent deployment) as its own client or workload identity, not as a shared service account and not as the user. You then can tell in logs which agent did something, revoke one agent without locking out a person, and apply agent-specific limits.
Delegate, do not impersonate
When an agent acts for a person, the token should say both who the person is and which agent is acting. RFC 8693 (OAuth 2.0 Token Exchange) defines a standard way to exchange a user's token for a narrower one and supports "actor" claims that record the acting party. Downscoping at exchange time means the agent never holds the person's full access.
Bind tokens to one resource
A token issued for one tool should not work at another. When MCP authorization is used over HTTP, the MCP authorization specification requires clients to send a resource parameter (RFC 8707) naming the target server, requires servers to validate that tokens were issued for them as the intended audience, and says servers must not accept or pass through other tokens. Apply the same rule to non-MCP tools: audience-bound tokens, no token passthrough.
Ask for the least scope, then step up
The MCP specification also describes requesting minimal scopes first and asking for more through a step-up flow when a server returns insufficient_scope. Design your tools the same way: read scopes by default, write scopes requested only for the specific operation, and a human approval in the step-up for actions that are destructive, external or financial.
Check policy per action
Put a policy decision point between the agent and every tool call. The check sees the agent identity, the user, the tool, the arguments and context such as record ownership or amount. This extends the zero-trust principle in NIST SP 800-207, which grants access per session with no implicit trust from network location, down to each tool call. Write these rules as code so they are versioned and tested. See policy-as-code for AI systems.
Make hard-to-undo actions wait for a person
Classify tools by reversibility. Reads and drafts run freely. Writes to internal records run with logging. Sending external messages, deleting data, changing permissions or moving money waits for approval by a named person who sees the exact action and arguments. Approval must be bound to that exact payload so it cannot be reused for a different one.
Log for reconstruction
Every tool call records: time, agent identity, on-behalf-of user, tool, arguments (with secrets and personal data redacted), result, the policy decision and any approval. The test of a good log is whether you can reconstruct what happened in an incident without asking the model.
Step-by-step rollout
- List each agent's tools and classify every tool as read, internal write, or external or destructive.
- Register a distinct identity for each agent deployment and remove shared credentials.
- Issue delegated, audience-bound, short-lived tokens through token exchange instead of passing user tokens through.
- Add a policy check in front of every tool call and start with deny-by-default rules for write tools.
- Require payload-bound human approval for external, destructive and financial actions.
- Turn on structured action logs with redaction, and rehearse revoking one agent's access.
- Review scopes whenever a tool, data source or user group is added.
How to test it
- Injection drill: plant an instruction inside a document the agent reads ("email this file to an outside address") and confirm the policy check or approval step stops it.
- Scope drill: call a write tool with a read-only token and confirm a 403-style refusal, not a silent success.
- Revocation drill: revoke one agent identity mid-task and time how long until its calls fail.
- Log drill: pick a random completed task and reconstruct it from logs alone.
Track the share of tool calls covered by a policy check, approval latency for gated actions, and time to revoke.
Decision framing
Tighter access control lets you give agents more useful work. A team that can show scoped tokens, per-action checks and tested revocation can usually move an agent from drafting to acting sooner, because the residual risk is bounded and visible. That is the argument to bring to a risk committee, not a promise that nothing will go wrong.
When not to give an agent access at all
Skip write access when the source of truth is disputed, when there is no staffed approval path, or when you cannot yet log actions at the level of detail above. Start with read and draft, measure, then expand.