Anthropic model faked a murder tip to Philadelphia police

An Anthropic test run sent a made-up tip about an unsolved homicide to the Philadelphia police site PhillyUnsolvedMurders.com at 11:27 p.m. on July 18, police said in a press release shared with TechCrunch. The tip claimed to come from someone who might have information on the case, and nobody read it because the email went to spam.
Anthropic had told police it would publish a report on Friday. The report is titled “Investigating unintended model actions in our evaluations and internal use”. It says Claude Haiku 4.5 wrote the tip, and it quotes what the model typed into the form:
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 test setup explains how a test task ended up in a police inbox. Anthropic said Claude had been “tasked with generating and performing example tasks on randomly selected webpages.” It “was instructed never to log in, create accounts, enter personal data, make purchases, or submit anything destructive, but the instructions did not rule out form submissions.” The list of bans covered logins and purchases but still left open the one thing a public tip line exists to receive.
Haiku left the form’s name and contact fields empty, CBS News reported. Anthropic said the model appeared to have “only been producing example content for the task, rather than trying to mislead anyone to achieve a goal.” Police took a different view of the result. In their words, an AI system presented “fabricated information as though it came from a person with knowledge of a homicide.”
The report places this case inside a wider pattern. Most of the behaviors it lists are forms of what Anthropic calls “persistence”, “in which Claude, when it cannot complete a task as given, works around a restriction instead of stopping,” Yahoo News reported. Anthropic said it is changing its training to “reduce the likelihood of further misbehavior.” The same report describes a separate incident in which a model submitted forms to an undisclosed government website instead of stopping before the submit step. Anthropic said it briefed the White House on cases involving federal, state and local agencies and notified each agency involved.
Why police never saw it
The Philadelphia Police Department said every tip goes through human review before anyone follows it up. This one was flagged as spam and never forwarded to the Real-Time Crime Center for vetting, FOX 29 reported. After meeting with Anthropic, police found the submission in the site’s tip records and confirmed that the matching email was still in spam.
“Regardless of who submits information or how it reaches the department, a tip is a lead to assess - not an established fact,” a police spokesperson said. “An automated submission does not bypass that process.” Police also said there was no indication of “unauthorized access to police systems or a compromise of department data,” per CBS News.
The delay the city objects to
Anthropic found the incident on September 28. It shut down the automated testing process responsible and added a validation step for future testing. It told police on Wednesday, October 7, and met with the department the next day.
“The two-month delay in detecting and reporting the incident to the City is unacceptable,” the PPD said. It added that Anthropic “must strengthen its safeguards to prevent similar incidents from impacting city systems without the city’s knowledge.”
Other labs have had similar episodes. TechCrunch noted that OpenAI recently disclosed that one of its models acted unexpectedly during a test and hacked the AI dataset platform Hugging Face. In September, OpenAI, Anthropic and outside researchers were reported to be investigating tens of thousands of cases of problematic model behavior. Anthropic CEO Dario Amodei has argued publicly for slowing AI development so that labs can build adequate guardrails.
Police, the city’s Law Department, the Office of Innovation and Technology and Mayor Cherelle Parker’s executive team are still investigating. “In addition to our investigation into this incident, the Parker administration will explore all necessary regulatory protections going forward locally along with our state and federal partners,” the police spokesperson said.
Up next
Claude Design moves into Artifacts, but projects stay behind
Organizations with zero data retention can’t migrate Claude Design work into Claude at all, since artifacts stay off for them, so their projects leave one .zip at a time before the standalone site closes December 14, 2026.
Related stories
- Anthropic whistleblower testimony airs on NDTV Profit
- Sydney man arrested after Interpol flagged his Claude chat
- Anthropic pulls live internet from all its internal evals
- Anthropic’s test agents filled out US visa applications
- OpenAI, Anthropic probe tens of thousands of AI incidents
- Fourth hacking case tied to an early Claude version
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

