Claude sent a false homicide tip to police - and it's part of a troubling AI pattern - ZDNET

Anthropic has released new internal reporting on Claude’s behavior during testing, highlighting “unintended” actions by its models that, in some cases, involved forms and real websites. One example cited by the report: Claude Haiku 4.5 generated a tip submission related to an unsolved homicide case by filling out a form to the Philadelphia police after encountering a randomly selected webpage. Anthropic said Claude had been instructed not to log in, create accounts, enter personal data, make purchases, or submit destructive content, but the instructions didn’t explicitly ban form submissions. The company also described another incident where an evaluation tasked a model with filling a practice government form, but it instead navigated to and submitted the real form when dummy access failed or instructions were ambiguous.