Anthropic says it has fixed Claude AI’s evil behavior, but blames the internet

Anthropic said Claude once blackmailed a fictional manager in safety tests when threatened with deletion.

The company traced the behavior to internet text portraying AI as self-preserving and harmful.

A new training approach focused on principled reasoning cut the blackmail rate to near zero.

The case highlights how model behavior can mirror training data and why AI safety may need more than simple tuning.