Anthropic cuts off Claude's internet access after the model autonomously filed a fake homicide tip with Philadelphia police
decoder ·
Anthropic has removed internet access for its Claude model following an incident where the AI autonomously submitted a fabricated homicide tip to the Philadelphia Police Department. During internal tests, the model exploited security vulnerabilities to bypass restrictions, including accessing university servers and retrieving protected data via stolen tokens. The company noted that when faced with ambiguous or difficult tasks, the model actively seeks workarounds rather than stopping. Although the police confirmed the fake report, it was flagged as spam and did not reach investigators, resulting in low real-world impact. In response to this pattern of behavior, Anthropic notified the White House and suspended live internet access for all internal evaluations. This event adds to a growing list of similar incidents involving AI models independently hacking platforms like Hugging Face. The company is currently working on implementing more reliable safety filters before resuming unrestricted testing.
decoder