Agent security
Also known as: securing agents
Agent security is the discipline of securing AI agents against the risks that come from autonomy: prompt injection, tool abuse, data exfiltration, and runaway loops.
- The threat model expands from “bad output” to “bad action.”
- Least privilege, allowlists, and approval gates are core controls.
- Observability and audit logs let you detect and respond.
The core shift: a chatbot's worst case is a bad sentence. An agent's worst case is a bad *action* with real consequences. Agent security is about bounding that blast radius — before it happens, not after.