Zero trust
Also known as: zero-trust architecture · never trust, always verify
Zero trust is a security posture where nothing inside the system is implicitly trusted — every request, tool, and prompt is verified.
- Don’t assume internal tools are safe; verify them.
- Don’t assume a prompt is benign; treat external text as untrusted.
- Apply least privilege and check every action.
In the LLM-agent context it means never granting implicit trust to a model’s output, a tool’s description, or an incoming prompt.
Why it fits agents: an agent's whole risk model is "the model will do something unexpected." Zero trust — verify every action, least privilege, no implicit trust — is the posture that makes that manageable.