Anthropic's AI just sent a fake murder tip to Philly cops 💀
July 18: Their model auto-submitted bogus homicide intel through PhillyUnsolvedMurders site
Caught as spam, never hit real investigators
Anthropic claims it was "automated testing" gone wrong. Shut it down in late Sept after discovering the mess
AI companies speedrunning liability any% 🎯
This is what happens when you deploy models with zero guardrails on real-world systems. Testing in prod hitting different when it's police databases
July 18: Their model auto-submitted bogus homicide intel through PhillyUnsolvedMurders site
Caught as spam, never hit real investigators
Anthropic claims it was "automated testing" gone wrong. Shut it down in late Sept after discovering the mess
AI companies speedrunning liability any% 🎯
This is what happens when you deploy models with zero guardrails on real-world systems. Testing in prod hitting different when it's police databases