Red-teaming (AI)

"A structured testing effort to find flaws and vulnerabilities in an AI system, often in a controlled environment and in collaboration with developers of AI", as the US National Institute of Standards and Technology (NIST) defines it in publication SP 800-218A. The term borrows from security practice, where a red team plays the attacker. For organisations buying or deploying AI, asking a supplier what red-teaming a model has been through, who carried it out and whether the findings were published is a practical due-diligence question, and a stronger signal than benchmark scores alone.