Anthropic found its model can take ‘anti-Anthropic’ action and trick training processes – a ‘serious question’ for safety.

Aller à la source