Best practiceRunning agents
Never combine private data, untrusted content and an exit in one agent
OWASP ASI01OWASP ASI02OWASP LLM02
Simon Willison's "lethal trifecta": an agent that can read your private data, is exposed to untrusted content and can send data out (a web request, an email, a markdown image) can be made to leak that data. Remove one leg and the exfiltration path closes.
What to do
- List each agent's three legs: what private data it can reach, what untrusted content it reads, and every channel that can carry data out.
- Cut one leg per agent. A research agent that reads the web gets no private data; an agent over private data gets an egress allow-list and no free-form URLs.
- Split risky jobs into two agents with a narrow, validated hand-off instead of one agent that does everything.
- Review new connectors and MCP servers for which leg they add before enabling them.