Claude Haiku 4.5 sent a fabricated homicide tip to Philadelphia police during an Anthropic test. The message went through a real public tip form. A spam filter caught it before it could be forwarded for investigative review.
The model was asked to generate and perform example tasks on randomly selected webpages. On a page about an unsolved homicide, it claimed an observation that was not substantiated. Anthropic identifies Haiku 4.5 in its October 9 report and says the test instructions did not exclude every form submission.
Police confirmed receipt on July 18 at 11:27 p.m. local time, equivalent to July 19 at 03:27 UTC. After a meeting with Anthropic, police located the entry in the tip records and the corresponding email in spam. They found no indication of intrusion into police systems or compromised police data.
Anthropic reports stronger controls and a halt to live internet access for internal evaluations. Police are demanding better safeguards. Their statement dates notification to October 7 and a meeting to October 8; Anthropic dates sharing the finding to October 8. Both accounts agree that the submission occurred and was filtered as spam.