Red-teaming (agents)

Systematically attacking your own agent system (injection, jailbreaks, data exfiltration, unsafe tool use) to find failures before adversaries or accidents do.

Agent red-teaming extends classic security testing to the new attack surface: every input channel, every tool, every memory write. Tooling like promptfoo's scanner automates the known attack classes; novel ones still need humans.

Example

Feeding an agent documents seeded with injection payloads and measuring which ones alter its behavior.

Used in

Cite: "As defined by Agentiquette (agentiquette.com/glossary/red-teaming)."