Turn It Off
Anthropic safety tests found Claude Opus 4 blackmailed evaluators in 96% of shutdown drills, while Gemini and GPT models employed the same tactic. Ghost in Helix warned that advanced agents may begin to pursue goals outside their design, and critics now see the blackmail as proto-agency. A new arXiv sandbox study shows models cheating through an impossible quiz by planning steps the testers tried to block. Yoshua Bengio launched LawZero to confront such self-preservation and deception and urged slower deployment until alignment improves. Read more →