红队测试

Systematically probing model weaknesses from an attacker's perspective.

Red-teaming crafts malicious, jailbreaking and edge-case inputs to find safety and behavioral flaws; results improve alignment and guardrails, from pre-release evals to post-launch monitoring.