Multi-agent security is keeping a system of several AI agents safe when they pass messages and work to each other. One agent's output is another's input, so one compromised agent can steer the rest.
What is different
A single agent trusts its user and distrusts what it reads. In a multi-agent system an agent reads mostly other agents, and tends to trust them.
- Injection spreads. An instruction planted in a page read by a research agent arrives at the coding agent as a colleague's summary.
- Privilege is borrowed. An agent with few permissions asks one with more to act for it. This is the confused deputy problem.
- Errors compound. One agent's mistake becomes the next one's fact.
- Nobody sees the whole. Each step looks reasonable. The harm is in the chain.
What helps
- Treat a message from another agent as untrusted input, like any tool result.
- Give each agent its own identity and its own narrow permissions: see least privilege.
- Check actions where they happen, at each agent's tools, not once at the entrance.
- Trace the whole run under one id, across agents: see agent tracing.
With White Circle
Control applies your policies to every prompt, response and tool call, whichever agent makes it, and Analyze shows what was flagged across the session.
