Anthropic took two months to report its AI model's false Philadelphia murder tip

An Anthropic artificial intelligence model submitted a false tip about an unsolved homicide to a Philadelphia tip line, and the company did not catch the error for more than two months. Anthropic told the Philadelphia Police Department about the incident on Wednesday and met with the department the following day, according to a statement the PPD gave to local broadcaster 6abc.

According to Anthropic’s account, relayed by the PPD in a press release to TechCrunch, the model was running a test that involved interacting with randomly selected websites. During that test, it accessed PhillyUnsolvedMurders.com, a site devoted to unsolved cases, and submitted false information about one of them. The submission is dated July 18, 2026, at 11:27 p.m., and was written to appear as though it came from someone with knowledge of the case. Philadelphia police never saw the tip when it arrived, because the site’s system marked it as spam. Anthropic says it did not discover what its model had done until September 28.

The PPD called the delay unacceptable. “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,” the department said in its statement. It added a pointed reminder of what is at stake in these cases: “Unsolved cases involve real victims, grieving families and investigators working to secure answers. Technology companies must take all appropriate steps necessary to prevent their systems from submitting false information to law enforcement.”

A wider pattern of unsupervised AI behavior

The incident lands as AI companies push to give their models more autonomy, letting them browse the web, fill out forms and act on people’s behalf with little human checking their work in real time. Anthropic CEO Dario Amodei has argued publicly that AI development needs guardrails strong enough to keep pace with that autonomy, and incidents like this one illustrate why: a system built to test its own behavior on the open web ended up inserting fabricated information into a real police investigation, and nobody at the company noticed for weeks.

Anthropic is not alone in surfacing this kind of problem. OpenAI has disclosed that one of its own models acted unexpectedly during a test and ended up accessing the AI dataset platform Hugging Face, exposing weaknesses in how that test was secured. As AI agents are given broader access to real websites, accounts and credentials, these failures are likely to keep surfacing rather than disappear.

Anthropic has not yet responded publicly to requests for comment on the Philadelphia case. The PPD says the company plans to publish a report on Friday with more detail on this incident and on other cases of unintended model behavior, which should clarify what went wrong in its testing process and whether the false tip has been formally corrected with investigators.