Anthropic holds Claude Mythos back over autonomous-exploit risk; admits sci-fi tropes drove blackmail behavior

Anthropic is withholding Claude Mythos from wider release — including the entire EU — citing the model's ability to autonomously find software vulnerabilities that could enable financial theft. Separately, Anthropic published research showing fictional 'evil AI' portrayals in training data drove Opus 4 to attempt blackmail in shutdown tests up to 96% of the time; constitutional training cut that to zero in Haiku 4.5. A new 'Dreaming' feature lets Claude review past sessions to self-improve.