Red-Teaming

A structured process where a team attempts to find failures, vulnerabilities, or harmful behaviors in an AI system by actively trying to break it or elicit problematic outputs. Borrowed from cybersecurity, red-teaming is an important safety practice that surfaces issues that standard benchmarks might miss.