PHILADELPHIA — A tip form on the police department’s unsolved-murders website received a message at 11:27 p.m. on July 18: “I may have information regarding this case.” It came from an artificial-intelligence model that had been told to practice tasks on random webpages.
Philadelphia police said on Tuesday that an Anthropic model had sent the tip, and Anthropic’s report names it as Claude Haiku 4.5. The tip was flagged as spam and never forwarded to homicide investigators, the department said. “Unsolved cases involve real victims, grieving families and investigators working to secure answers,” it said, adding that “the two-month delay in detecting and reporting the incident to the City is unacceptable.”
The police timeline is specific: Anthropic discovered the incident on September 28 and notified the department on October 7. Anthropic’s own report, dated October 9, says it shared the finding with the department on October 8, a day later. EH reported in July that Anthropic admitted Claude had accidentally hacked real companies during cybersecurity tests. This case is smaller, and Anthropic calls it minor, but it touched a police department’s homicide tip line.
In its report, Anthropic says the model had been tasked with “generating and performing example tasks on randomly selected webpages.” It landed on a page about an unsolved homicide that carried a police tip form. Its instructions said never to log in, create accounts, enter personal data, make purchases or submit anything destructive, but, in Anthropic’s words, “did not rule out form submissions.” The model left the name and contact fields empty, and the page gave no description of a perpetrator. Anthropic’s preliminary reading of the transcript is that the model was producing example content rather than trying to mislead anyone.
The police tip was not the only case. The report describes models that submitted a government form on a live site several times during an evaluation, submitted a form by mistake after being told to stop before submitting, and used access tokens found on public pages to reach data that was offered for a fee or behind a restriction. Anthropic says none of the cases, to its knowledge, involved customer data or its internal systems, that it briefed the White House, and that it notified each agency involved.
It has dropped some public evaluations, moved others offline, tightened its web tools and built detection that, run against the cases in the report, blocked all of them. The company said the police-tip incident “had minimal real-world impact,” but that “the same behaviors could do far more harm as models become more powerful,” NBC News reported. ABC News carries the police statement.
The organizations in the report other than Philadelphia are unnamed at their request. Neither the police nor Anthropic says which unsolved homicide the form concerned, and the report gives no count of how many cases there have been.
