Red-teaming is the only reliable way to verify AI safety measures, researcher argues

dfrsrchtwts · x · 2026-11-11

AI safety researcher dfrsrchtwts argues that red-teaming is the only reliable way to check whether safety measures work, since there is no good science of AI motivations, capabilities, or control. A safety system that cannot be red-teamed therefore cannot be trusted.

In a follow-up he highlights a sharper problem: if safety training starts from initialization, red-teaming it requires an adversarial initialization rather than an adversarially trained model with bad motives — significantly raising the verification bar.

Original post →

Red-teaming is the only reliable way to verify AI safety measures, researcher argues

More from Safety