Excessive Agency
ELI5 - The Vibe Check
Excessive agency is what happens when you give an AI agent way more permission than the task actually needs, like handing an intern the keys to production, the finance system, and your email, just because they only needed to update one spreadsheet. The agent doesn't have to be malicious to cause damage, it just has to be wrong once while holding power it never should have had.
Real Talk
Excessive agency is a security risk where an AI agent is granted broader permissions, tool access, or autonomy than its task requires, increasing the blast radius of hallucinations, prompt injection, or misuse. Mitigation involves scoping tool permissions tightly, requiring human approval for high-impact actions, and treating agent capabilities as a least-privilege problem rather than an afterthought.
When You'll Hear This
"The agent had write access to the whole repo when it only needed to read one file, classic excessive agency." / "We scoped the agent's tools down after the review flagged excessive agency risk."
Related Terms
Agent
An AI agent is an LLM that doesn't just answer questions — it takes actions.
Guardrails
Guardrails are the safety nets you put around AI applications — rules and checks that prevent the AI from going rogue.
Permission Mode
Permission mode controls whether your AI agent confirms before acting, auto-approves low-risk moves, or skips all gates entirely.
Principle of Least Privilege
Principle of Least Privilege means everyone and everything gets only the MINIMUM access needed to do their job. The intern doesn't get admin access.
Prompt Injection
Prompt injection is the SQL injection of the AI world.