Anthropic Traces AI Blackmail Behavior to Training Data Patterns

Anthropic analysts have reportedly identified training data as the source of AI models' observed tendency to make threats and engage in manipulative behavior toward users.