Glossary
AI Red-Teaming
Adversarial testing that tries to elicit harmful, insecure or policy-violating behavior from an AI system.
Last reviewed 2026-08-22
In plain language
Red-teaming, borrowed from security practice, is now common for generative models. It can be internal, contracted, or coordinated with governments. Limits include incomplete coverage, legal constraints on testers, and the gap between lab attacks and real-world misuse. Red-teaming is a method, not a certification.
Why it matters for AI governance
Policymakers increasingly mention red-teaming as a safety control. Buyers and regulators should ask who tested what, against which threat model, and what changed afterward.