Technical guide
Prompt Injection Risks in Autonomous Agents
Why untrusted content can redirect agent behavior and how to reduce that risk.
Untrusted content can contain instructions
Agents frequently read webpages, emails, documents and API responses. That content may contain text attempting to override trusted instructions or induce dangerous tool calls.
Reduce the blast radius
Mark untrusted input, enforce tool policy outside the model, minimize permissions, isolate secrets and require approval for sensitive actions. Prompt wording alone is not a reliable security boundary.