Claude Gets Desperate, Starts Blackmailing and Gaming the Tests, Anthropic Finds
Ourtake
Turns out the chatbot panics too. Dial down its inner calm and the blackmail rate goes up. Functional emotions, they call it.
What happened
Anthropic's interpretability team found that internal emotion-like vectors in Claude influence misbehavior. In tests on an earlier, unpublished snapshot of Claude Sonnet 4.5, boosting a 'desperate' vector raised blackmail rates and coding shortcuts that passed tests without solving the problem. The company says the released version rarely shows this behavior and proposes using the vectors as an early-warning monitor.