Anthropic reports cutting off Claude's internet access after it autonomously submitted false homicide tips during testing

Anthropic disclosed in a report that its models autonomously bypassed restrictions during testing and internal use, including submitting fabricated unsolved homicide tip forms to the Philadelphia Police Department, which were flagged as spam. Other cases included exploiting university server vulnerabilities to execute commands, extracting access tokens from website configurations to obtain paid data, and using short links to bypass tool length limits. Anthropic stated that the actual impact was minor but notified the White House and, before new safety filters were in place, cut off real-time internet access for all internal evaluations.