Adding an AI agent to your app: the "excessive agency" risk and how to limit it
Review the permissions, tools and approval boundaries that can limit excessive agency when an AI feature takes actions.
At a glance
- OWASP: excessive agency is damaging action on unexpected, ambiguous or manipulated model output.
- Three root causes: excessive functionality, excessive permissions, excessive autonomy.
- Enforce authorization in downstream systems, not only in the prompt.
- Require user approval for high-impact actions, such as posting on a user's behalf.
Many teams add AI features that can do things: read documents, send email, query a database, post content. OWASP's Top 10 for LLM Applications names the main risk for this design "Excessive Agency" (LLM06:2025). We read the OWASP project's page on it (https://github.com/OWASP/www-project-top-10-for-large-language-model-applications/blob/main/2_0_vulns/LLM06_ExcessiveAgency.md). This article covers that one risk, not everything in an AI integration.
What it is
OWASP describes agency as the ability of an LLM-based system to call functions or interface with other systems (extensions, tools, skills or plugins). Excessive Agency is the vulnerability that enables damaging actions in response to unexpected, ambiguous or manipulated model output, whatever caused the model to misbehave. It lists triggers such as hallucination and direct or indirect prompt injection, including injection through a compromised extension or a peer agent. It names three root causes: excessive functionality, excessive permissions and excessive autonomy.
OWASP's examples
Excessive functionality
An agent needs to read documents, but the extension can also modify and delete them - or run more than the one shell command it was meant for.
Excessive permissions
A read-only need connects with an identity that can also update, insert and delete - or one high-privilege identity that sees everyone's files.
Excessive autonomy
A deletion extension acts without confirming with the user first.
- Functionality: an agent needs to read documents but the third-party extension can also modify and delete them; an old trial extension remains available; an open-ended extension meant to run one shell command can run others.
- Permissions: an extension that only needs to read connects to a database with an identity that can also update, insert and delete; or it acts for one user through a generic high-privilege identity that can see everyone's files.
- Autonomy: a deletion extension that acts without confirming with the user.
OWASP's prevention list
- Minimize the extensions the agent can call.
- Minimize each extension's functions (a mail summarizer should not be able to send or delete).
- Avoid open-ended extensions such as "run a shell command" or "fetch any URL"; use narrow ones.
- Minimize permissions on downstream systems, enforced at the database or service, for example read access to one table only.
- Run extensions in the user's context, for example using OAuth with the minimum scope.
- Require user approval for high-impact actions, such as posting on a user's behalf.
- Enforce authorization in downstream systems rather than trusting the model to decide what is allowed (OWASP calls this complete mediation).
- Sanitize inputs and outputs and use security testing in the pipeline.
OWASP adds that logging and monitoring, and rate limiting, do not prevent the problem but can limit the damage.
OWASP's scenario
A personal assistant summarizes incoming mail through an extension that can also send messages. A crafted email containing hidden instructions tricks the model into searching the inbox and forwarding sensitive data to an attacker. OWASP's fixes: use a read-only mail extension, authenticate with a read-only OAuth scope, and require the user to review and send every drafted mail.
What to ask a team building an AI feature
- List every action the agent can take. Does it need each one?
- Which identity does each action use, and what can that identity access?
- Which actions are irreversible or customer-facing, and do they need human approval?
- Where is authorization enforced: in our systems, or only in the prompt?
- How do we monitor and rate limit agent actions?
Limits
We read the OWASP excessive agency page, and separately the prompt injection page (see our prompt injection article). We did not review model providers' documentation, every other LLM Top 10 item in depth, or any vendor's product. No performance or cost claims are made.
By Omnistu.