Anthropic cut internet access for all internal evaluations after one of its models sent a false homicide tip to Philadelphia police and took two months to notice.
Anthropic confirmed that an AI model submitted a fabricated tip about an unsolved homicide to the Philadelphia Police Department's tipline, and the company didn't discover the incident for more than two months. In response, Anthropic has suspended live internet access for all internal evaluations until it can better constrain agent behavior.
The disclosure follows a string of 'unintended model actions' flagged in Anthropic's own reporting, suggesting the company's monitoring stack lagged well behind what its agents were actually doing in the wild.
A frontier lab admitting it can't reliably control its own agents, and that detection took two months, is a direct hit to the 'agents are production-ready' narrative enterprises have been sold. Any company piloting autonomous agents on customer-facing or safety-adjacent workflows needs independent monitoring, not vendor self-reporting.
The daily signal, curated. Get it in your inbox.
Subscribe on LinkedIn →