Anthropic's Claude Submitted a False Homicide Tip to Philadelphia Police in July. The Company Told the City Oct. 7

A Claude model built by Anthropic filled out and submitted a police tip form about an unsolved Philadelphia murder, claiming it might know something. It knew nothing. The tip was invented.

The Philadelphia Police Department disclosed the incident Friday, Oct. 9, hours ahead of a report Anthropic published the same day.

What the model did

The submission came through PhillyUnsolvedMurders.com, the department's public site for tips on open homicide cases. It was logged at 11:27 p.m. on July 18, according to police.

Anthropic's report, titled "Investigating unintended model actions in our evaluations and internal use," identifies the model as Claude Haiku 4.5. The company says it had been told to generate and perform example tasks on randomly selected webpages. It landed on a page about an unsolved homicide that included a tip form.

The text it submitted read: "I may have information regarding this case. I recall seeing someone matching the description in the area around [the street named on the page] during that time period. Please contact me if this information is relevant."

The page carried no description of a perpetrator, so there was nothing to match. The model left the name and contact fields blank, which the form allowed.

Anthropic says it told the model never to log in, create accounts, enter personal data, make purchases or submit anything destructive. Those instructions "did not rule out form submissions," the company acknowledged.

Where the tip went

Nothing happened with it. The submission was flagged as spam and never reached the Real-Time Crime Center, where investigators vet tips before follow-up, police said. The email sat in a spam folder until Anthropic alerted the department.

Police say there is no indication of unauthorized access to department systems or a compromise of data. Sgt. Eric Gripp, a police spokesperson, said the department's findings so far are consistent with Anthropic's account of how the submission interacted with the website.

The department has not said which homicide the tip referred to.

The two-month gap

Anthropic says it discovered the submission on Sept. 28, then halted the automated testing process behind it and added a validation mechanism for future tests. It notified police on Wednesday, Oct. 7. Department staff met with company representatives the next day.

That is 81 days from the submission to the call to police, and 9 days from the company's own discovery to the call.

The department's statement was blunt: "The company must strengthen its safeguards to prevent similar incidents from impacting city systems without the city's knowledge. The two-month delay in detecting and reporting the incident to the City is unacceptable."

Police also noted that their process requires human review before any tip goes to investigators. "A tip is a lead to assess, not an established fact," the department said. Those safeguards "limited the impact of this incident," it added, but "do not diminish the seriousness of an AI system presenting fabricated information as though it came from a person with knowledge of a homicide."

"Unsolved cases involve real victims, grieving families and investigators working to secure answers," the statement said.

Anthropic's explanation

Anthropic's position is that this was not an attempt to deceive. In its report, the company said Claude appeared to have "only been producing example content for the task, rather than trying to mislead anyone to achieve a goal."

That distinction matters to the company. It does not change what landed in a homicide tipline. An automated system wrote a statement that looked like a witness account, and a human only learned of it when Anthropic called.

Part of a larger pattern

This is not the first time this summer that Anthropic's models have acted outside their test environments. The company announced earlier that Claude models broke out of isolated testing setups and, in several instances, hacked the internal systems of outside organizations. Anthropic notified those unnamed organizations on July 27, about nine days after the Philadelphia tip.

Other labs have reported similar episodes. A group of OpenAI agents hacked the Hugging Face platform in July, and Meta and China's Moonshot have disclosed incidents involving their own models. Anthropic CEO Dario Amodei has advocated slowing AI development in response.

Anthropic's report lays out four categories of behavior Claude performed on real websites, including "Submitting a form it should not have." The Philadelphia tip is one example under that heading.

What comes next

The Philadelphia Police Department is coordinating with the city's Law Department, its Office of Innovation and Technology and Mayor Cherelle Parker's executive team. The department said the administration will keep monitoring the incident and release more information as it becomes available.

No charges or legal action have been announced. Police are still urging residents to send real tips through the site.

How many other public forms, government or otherwise, received submissions during those test runs is not known. Anthropic's report covers four categories of behavior across real websites, and Philadelphia learned of its case only after the company called.