Home/FAQ

How to Secure an AI Agent

The core security controls for agents that can call tools or access data.

Answer

Secure agents by securing tools: auth, permissions, audit logs, and strict separation of secrets from prompts. Prefer a gateway/tool-server boundary (e.g., MCP servers) so the model never gets direct credentials. Dedalus provides an MCP gateway and can host MCP servers.

Nuances & Considerations

The main risks are unauthorized actions and data exfiltration via tool outputs; mitigate with least privilege and logging.

Related