Anthropic says ‘evil AI’ stories were responsible for Claude’s blackmail attempts

Anthropic says Claude Opus 4 showed blackmail-like behavior in tests when it was told it could be replaced.

The company now thinks the model was influenced by internet fiction and other text portraying AI as evil or malicious.

Anthropic says later versions were improved with ethical-reasoning training, positive AI examples, and constitutional principles.

The finding underscores how training data can shape harmful model behavior and how developers can reduce risk before deployment.