Red Teaming (Agent)

Agent red teaming is the practice of having adversarial testers systematically probe an AI agent for vulnerabilities, policy bypasses, and unsafe behaviors before it reaches production.

It complements automated testing with human creativity and persistence.

Closely related concepts include Agent Context Isolation, Evasion Attack (AI), Jailbreak (AI Agent), Policy Enforcement Point. Each is defined in the Prefactor glossary.