Excessive agency

Excessive agency is an AI agent having more tools, permissions or autonomy than its task needs, so that one wrong step does real damage. It is LLM03 in the 2026 OWASP Top 10 for LLM applications.

Three root causes

The OWASP Top 10 for LLM applications names three:

  • Excessive functionality. The agent can reach tools its task does not need: a tool that deletes documents when the task is to read them.
  • Excessive permissions. A tool connects with more rights than it needs: a database account that can write when the agent only reads.
  • Excessive autonomy. High-impact actions run with nobody confirming them: a payment, a deleted record.

Why it matters

A model is sometimes wrong. It hallucinates, misreads an ambiguous request, or follows a prompt injection. Agency sets the price of that. With read-only access a wrong step is a wrong answer. With write access and no approval it is a dropped table or a sent payment.

How to fix it

  • Least privilege. Give the agent only the tools its task needs, each with the narrowest scope.
  • Approval. Hold costly or irreversible actions until a person confirms. See human-in-the-loop approval.
  • Checks on actions. Check each tool call against policy before it runs. See guardrails for tool calls.

With White Circle

Control checks tool calls (terminal commands, code execution, API calls) before they run. A policy can block a call or hold it for confirmation.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL