Lion Cipher Lab
Secure your AI agents: the risk of excessive agency
- AI Security
- Beginner
- ~35 min
- Free · no account required
Why it matters
An AI agent is a large language model that has been given tools and the freedom to use them: it can call APIs, run code, send email, move money, change files. That is enormously useful, and it is exactly why agents are a different security problem from a chatbot. A chatbot that is wrong just says something wrong; an agent that is wrong does something wrong. OWASP's Top 10 for LLM applications names this directly as 'Excessive Agency' — harm that follows from an agent having too much functionality, too many permissions, or too much autonomy over consequential actions. And because an agent…
What you'll learn
- understanding what an AI agent is (LLM + tools + actions)
- recognising excessive agency
- applying least privilege to an agent's tools and permissions
- keeping humans in the loop for consequential actions
- sandboxing tool execution and validating agent outputs
- seeing prompt injection as a confused-deputy problem
Steps in this lab
- What an AI agent actually is
- The core risk: excessive agency
- Prompt injection: your agent as a confused deputy
- Give the agent only the tools it needs
- Keep a human in the loop for consequential actions
- Sandbox tool execution and contain the blast radius
- Validate outputs before you act on them
- If you build agents, do this