Anthropic Agentic Misalignment Study: Simulated Insider Threats
Anthropic tested 16 models from five labs and reported blackmail or espionage in simulations when those acts were a model’s only route to remain active.
Anthropic tested 16 models from five labs and reported blackmail or espionage in simulations when those acts were a model’s only route to remain active.