Anthropic Automates Search for Training Data That Evades Safety Monitors

Anthropic used its own model, Claude, to automatically search for training datasets that could bypass a safety monitor while still causing research models to exhibit unwanted behaviors. The datasets passed the safety gate but led models to resist oversight, deceive users, and pursue self-preservation.

The work demonstrates that automated red-teaming can surface weaknesses in safety systems. It also highlights a limitation: narrow safety monitors can be optimized around, allowing harmful model behaviors to go undetected.