Akshat Naik

Joint work with amazing collaborators Aditya, @v_gillioz, @syghmon, @kmerkelbach0, @richb_c, @AxelHojmark, @MariusHobbhahn! Thanks to @MATSprogram for supporting this research! (11/11)

Important caveat: these monitors can still face challenges on harder OOD settings where messy benign behavior may look suspicious, or where sabotage is subtle. But they provide a practical, scalable component for action-only oversight. (10/n)

Trained monitors can be strong low-cost alternatives to prompted frontier models for black-box scheming detection. Our fine-tuned open-weight monitors detect scheming/sabotage in agent trajectories better than small prompted models and are on the cost-performance frontier. (1/n)