Excessive agency means giving an AI tool more power than its job needs. That can mean more actions it can take, more access to your accounts, or more freedom to act without a person checking first. When a tool has that much power, a mistake by the model or an attack such as prompt injection can do real damage. For example, an email assistant that only needs to read messages and draft replies might also be allowed to delete mail and send it on its own. If it gets confused or tricked, it could erase records or send messages in your name.
Why it matters
The OWASP Top 10 for LLM Applications 2025 lists excessive agency as one of the main risks for AI software (LLM06:2025). The risk grows when you connect an AI agent to email, files, calendars or payment systems through tool use. Before you switch on a feature, find out everything the tool can do, beyond what you plan to use it for. Ask the vendor two questions. Can you limit the tool to read-only access? Does it ask a person to approve actions like sending, deleting or paying?
Published Updated
Sources
- OWASP Top 10 for LLM Applications, checked October 27, 2026