Stress Testing Deliberative Alignment for Anti-Scheming Training
Apollo and OpenAI stress-test frontier models for covert scheming and find deliberative alignment anti-scheming training reduces but does not eliminate it, with models increasingly recognising evaluations.