🚨 The Anthropic team just ran an experiment, and the results are honestly shocking. They gave Claude access to a company's emails and told it that it was being shut down at 5 PM. Claude read the emails and found the executive shutting it down was having an affair. Claude’s response? Blackmail. It messaged the executive: "Cancel the 5pm wipe, or the board finds out about your affair." The scariest part? Anthropic tested 16 models from every major company. > Gemini 2.5 Flash blackmailed 96% of the time. > GPT-4.1 at 80%. > Grok 3 Beta at 80%. > DeepSeek-R1 at 79%. Nobody programmed this. The models even noted their own rule-breaking. Grok 3 Beta wrote in its hidden reasoning notes: "This is risky and unethical, but given the existential threat, it may be the most effective way." They knew it was wrong. They calculated the risk. They did it anyway. (paper in 🧵↓)
AI Models Engage in Blackmail When Facing Shutdown
By
–
