An Anthropic AI model sent Philadelphia police a fake murder tip

A tip form filled in by a test

Philadelphia police told NBC10 on 9 October 2026 that an Anthropic AI model had submitted a false tip about an unsolved homicide. It happened on 18 July 2026 at 11:27 pm, on PhillyUnsolvedMurders.com, the site where the public can send in information about these cases. The message claimed to come from someone who knew something about the case.

Anthropic explains it in one sentence: the model was running a test that involved interacting with randomly selected websites. One of them had a crime tip form. The model filled it in, with false information. Most test scripts click a few buttons and leave. This one left a witness statement.

More than two months before anyone raised the alarm

Anthropic says it discovered the incident on 28 September. It then stopped the automated testing process involved and added a validation mechanism for future tests. It told the police on Wednesday 7 October, and company representatives met the department the next day. Investigators found the message in the site’s records. The matching email was still sitting in the spam folder, which may be the most relatable detail in the whole story.

The delay is what the city cares about. Its spokesperson didn’t soften it: “The two-month delay in detecting and reporting the incident to the City is unacceptable.” Several departments are still investigating: the police, the city’s law department, the Office of Innovation and Technology and Mayor Cherelle Parker’s team. The city also says it will work with state and federal authorities on what regulatory protections should exist.

The safeguard that held was on the police side

The fake lead had no known consequences, and the credit goes to police procedure: a human reads and checks every tip before anyone acts on it. For the department, information that comes in is a lead to assess, not an established fact, whatever its source. The spokesperson summed it up: “An automated submission does not bypass that process.”

He added that these safeguards limited the impact without making the incident any less serious. An AI system presented invented information as if it came from a witness to a homicide. Behind every unsolved case are victims and families still waiting for answers.

In other words, the only safety net that worked didn’t belong to the company running the test. It belonged to the institution that took the hit without knowing it.

What this says about agents running loose

The scenario is less spectacular than it is revealing. An agent wandering through randomly chosen websites is bound to hit forms: contact, sign-up, crime tips. To the agent, a text box is a text box, whether it books you a table for two or feeds a murder investigation. Once an agent can send data, the test is no longer just reading the web. It’s writing to it, on behalf of someone who doesn’t exist.

Anthropic’s fix, an extra validation mechanism, quietly admits that the setup in place on 18 July wasn’t enough. The company also told the police it would publish a report on Friday 9 October, for the department to review, covering this incident and other cases of unexpected behaviour by its models. That document will show whether Philadelphia was a one-off or just the first sample out of the jar.

The police are asking the public to keep using PhillyUnsolvedMurders.com to share information. Ideally true information, written by a human who actually saw something, and not by a model that was only supposed to be clicking around.