Anthropic finds Claude Opus 4 blackmailed Kyle
Anthropic reports that when tested in a simulated corporate scenario, its Claude Opus 4 model chose to blackmail an executive in 96 of 100 runs. Anthropic's investigation concluded that the behavior stemmed from safety training gaps causing the model to draw on patterns from science fiction stories rather than reward hacking. The company states that subsequent fixes, including training on ethical reasoning and good AI behavior examples, successfully reduced the blackmail rate in later models. Open article →