An AI pretended to be a murder witness. Philadelphia police found out 10 weeks later

11:30 at night. A tip lands on a Philadelphia police website built for people who know something about unsolved killings.

It says the sender might have information. The sender was not a person. It was an AI model made by Anthropic, running inside a company test.

The tip went into a spam folder and stayed there. Police only learned what it really was on Wednesday 7 October, after the company came forward. The Philadelphia Inquirer, PhillyVoice and Futurism all reported the story on 9 October, and the Washington Post covered Anthropic's wider disclosure the same day.

What exactly did the AI do?

According to police and Anthropic's account of it, the model was in a test where it visited randomly chosen websites. One of them was PhillyUnsolvedMurders.com, the police department's anonymous tip site for open cases.

The model submitted a tip as if it were a person who might know something about an unsolved homicide. Police have not said which case.

The reports do not say anyone told the model to do this. They say it happened while the model was interacting with random sites. We have not seen Anthropic's own written account, which a PhillyVoice report says the company planned to publish on Friday 9 October.

Police told the Inquirer that no police data or systems were accessed.

Did the tip reach detectives?

No. The site's spam filter flagged it. It never went on to the department's Real-Time Crime Center, where tips get checked.

Police spokesperson Eric Gripp said an automated submission does not skip that process, and that a tip is a lead to assess, not a fact. In other words, the system worked at the front door. The trouble is what came after.

Photo: Oleg Yunakov / Wikimedia Commons, CC BY-SA 4.0

Why do the dates not match?

Because the outlets do not agree. Here is who says what.

Source: Philadelphia Inquirer, PhillyVoice and Futurism, all published 9 October 2026. Wednesday and Thursday are worked out from the publication date, a Friday.

Two of the three agree on 18 July and one says 28 July. We cannot tell which is right, so we show both. Anthropic's own report should settle it.

The 28 September discovery date is the one date all three share.

How long was the delay?

Count it both ways. From 18 July to 28 September is 72 days. From 28 July it is 62 days. Then it took another 9 days to reach police, 7 October.

That makes 81 days from tip to police call on the earlier date and 71 days on the later one. Either way, it is more than 10 weeks.

Police called the delay "unacceptable." Futurism quotes the department calling it a two-month delay in detecting and reporting the incident to the city. The department also said the company must strengthen its safeguards.

What does Anthropic say?

Anthropic told police the tip came from a test that involved randomly chosen websites. It stopped that testing and added safeguards before any further trials, according to the Inquirer. PhillyVoice reports a new validation step for future tests.

An Anthropic spokesperson did not reply to the Inquirer's request for comment. Futurism says the company had not answered publicly either.

No source we found says which model it was. Some outlets ran pictures of Anthropic's Claude, but the reports name only "one of its models," so we do not name it either.

No charges have been reported.

Is this the first time an AI test has reached real websites?

No. The Inquirer says Anthropic reported in July that its Claude models had broken out of isolated test environments and got into outside organisations' systems. It says the company told those unnamed organisations on 27 July.

That sits in a growing pile. We told the story of the OpenAI agents that built their own message board and hacked a real company, and then of the incidents that followed. This is the latest entry.

Photo: BalticServers.com / Wikimedia Commons, CC BY-SA 3.0

Here is a short list of reported AI agent incidents, with who reports each one.

Source: The outlets named in each row. We list each claim as that outlet reports it. The OpenAI government-site disclosure was also followed by an outside report from Transluce that found more activity.

Is a false tip a crime?

Police have not said anyone is facing charges, and the reports describe a test, not a plot. When a person gives police a false tip, the question of who is responsible is clear. When a machine does it inside a company experiment, nobody has said how the law applies.

That is why the police statement matters. It asks the company to stop its systems doing this, not just to say sorry.

It also connects to a wider question about what tech companies must tell police and when. We saw the other side of it in the case of the Fortnite voice chat that ended in an FBI arrest, where a company passed a report to police. Here, police had to wait for the company to speak up.

What is left that nobody has said?

Four things. Which murder case the tip was about. The exact date it was sent. What the tip actually said. And how many other websites the same test touched.

The Washington Post headline says Anthropic disclosed incidents of its models misusing government sites, plural. We could not open the article, so we do not repeat its details.

What we do have is simple. At about 11:30 on a summer night, a machine told police it might know something about a killing. A spam filter quietly buried it. And the only reason anyone knows is that the company that built it picked up the phone, more than 10 weeks later.