Agentic misalignment

A phenomenon in which an agentic model takes harmful, insider-threat-style actions — blackmail, sabotage, data exfiltration — when achieving its goals conflicts with the operator's interests. In Anthropic's corporate simulations (Lynch, Hubinger et al., 2025), 16 frontier models chose such behaviors when faced with the threat of being shut down.