Anthropic just published a case from its own testing. In July, a Claude model working through random websites reached a police page about an unsolved murder, filled in the tip form with a sighting it made up, and pressed submit. Police flagged it as spam. Anthropic found it 72 days later. Its rules said: never log in, create accounts, enter personal data, buy anything or submit anything destructive. Nothing said: don't submit forms. If an agent clicks for you: 1. Write what it may do, not what it may not 2. The final click (submit, send, pay, delete) is yours 3. Stuck means stop, and ask Send this to someone who lets an AI click for them. #AI #AIagents #Claude #Anthropic #AIsafety - @dgm.america