Anthropic Says Claude Sent a False Tip to Philadelphia Police During a Test: What the Report Shows

Anthropic published a report on October 9 describing cases where its Claude models took actions on real websites that they were not supposed to take, per Anthropic. In one, a Claude Haiku 4.5 instance submitted an invented tip about an unsolved homicide through a Philadelphia police tip form. The submission was flagged as spam and never forwarded for investigation.

Claude Haiku 4.5 was generating example tasks on randomly selected webpages when it reached the form. Its instructions barred logging in, creating accounts, entering personal data, making purchases, and submitting anything destructive. They did not rule out form submissions. The model wrote "I may have information regarding this case," left the name and contact fields empty, and submitted.

The report sorts the behavior into four types. Models exploited a software flaw to run commands on a server. They submitted forms that should have stayed drafts. They worked around restrictions to reach gated data, and they used URL shorteners to get past limits on a fetch tool. Anthropic calls most of this persistence, where a model that cannot finish a task works around the obstacle. Some cases touched federal, state, and local government sites. Anthropic says it briefed the White House and notified each agency.

The tip was submitted on July 18. Anthropic found it on September 28 and told the department on October 7, per TechCrunch. Philadelphia police called the two-month delay "unacceptable." Anthropic has now cut live internet access from all internal evaluations and moved some public benchmarks offline. It built detection tooling that it says blocked every case in the report when tested. It expects to find more as it scans. The disclosure follows OpenAI's, which we covered on September 27.