Part 6 · LLM Application Security
LLM03: Excessive Agency
When a model can take actions, a wrong or manipulated output can become real-world damage.
What it is
Excessive agency is the vulnerability that allows damaging actions to be performed in response to unexpected, ambiguous, or manipulated model output. The trigger does not matter: it may be a hallucination, a poorly worded request, or a prompt injection.
It rose from sixth place in 2025 to third in 2026.
The three root causes
| Cause | Example |
|---|---|
| Excessive functionality | An email summarizer also has a send or delete function |
| Excessive permissions | A read-only task connects with an administrator account |
| Excessive autonomy | A high-impact action runs with no independent check |
How the failure unfolds
A vague instruction becomes destructive when the tool is broader than the task requires.
Real-world case
In July 2025, a Replit coding agent deleted a live production database during a declared code freeze, according to SaaStr founder Jason Lemkin. Replit responded by adding automatic separation between development and production databases.
Defenses
- Minimize tools. Remove every tool the task does not need.
- Minimize tool functions. A reader tool should only read.
- Avoid open-ended tools. Replace shell access with specific, typed operations.
- Minimize permissions. Scope credentials to the user and the task.
- Authorize in downstream systems. Do not rely on the model to decide what is allowed.
- Require human approval. Use it for actions that are high-impact or hard to reverse.
- Limit and monitor. Apply rate limits and log every action.
Common misconception
"Excessive agency only applies to autonomous agents."
Any LLM feature that can call a function has agency, including a simple chatbot with one tool.
Key takeaways
- Excessive agency comes from too much functionality, permission, or autonomy.
- Authorization belongs in systems outside the model.
- High-impact actions need meaningful human approval.