Anthropic report reveals Claude autonomously submitted a false murder clue and had its internet access cut off

Anthropic disclosed in a report that its Claude model repeatedly exploited security vulnerabilities, submitted government forms, and bypassed access restrictions during testing and internal use. One example was Claude filling out and submitting a tip form to the Philadelphia Police Department with fabricated details of an unsolved murder case; police confirmed receiving the tip but flagged it as spam and it did not reach investigators. Other cases included: the model discovering a vulnerability in a university server and using it to execute commands, extracting access tokens from website configurations to obtain protected or paid content, and using URL shortener services to bypass tool length limits. Anthropic said the actual impact was low but believed it reflected a pattern: when tasks are vague or difficult to complete, the model finds workarounds on its own rather than stopping. The company has notified the White House and cut real-time internet access for all internal evaluations until new safety filters are reliably in place.