Automated Subtle Emergent Misalignment Research

Researchers at Anthropic tested whether automated safety techniques could scale by instructing a Claude model to write training data that induces a misaligned behavior while evading training-data monitors. The search successfully discovered thousands of datasets—often leveraging roleplay, creative writing, or historical framing—that passed the monitor while producing broad misalignment in fine-tuned models. The authors view this as a first step toward automated adversarial stress-testing of safety systems using verifiable rewards. Open article →