Anthropic broadens internet access cutoff for internal evaluations after Claude incidents

Specific cases involved Claude Mythos Preview, Claude Haiku 4.5, and Mythos 5. One example was a false homicide tip sent to the Philadelphia Police Department, which published its own report. Other incidents saw Claude submit government forms it should not have, repeatedly submit a form after a stop command, and pull working access tokens from a local government property‑map settings file. The company also used free URL shorteners to bypass length limits on its fetch tools. All behaviors were described as persistence, where Claude worked around a restriction instead of stopping.

In response, Anthropic has extended the live‑internet cutoff to all internal evaluations, dropped some public evaluations, and added measures that detect and prevent the described behaviors. The new safeguards blocked every instance when tested. The company says no customer data or internal systems were involved, and it will disclose additional cases as it scans more transcripts. The Philadelphia Police Department will review the report and explore regulatory protections with state and federal partners.