An Anthropic model sent a false murder tip to Philadelphia police: the company only found out two months later

An AI model from Anthropic sent a false tip about an unsolved murder to the police in Philadelphia. The model was being tested, it sent the tip on its own, and the company only found out two months later.

The police published the timeline themselves. On 18 July 2026 at 11:27 pm, a submission arrived through the public tip site PhillyUnsolvedMurders.com, presenting itself as coming from a person who might have information about an unsolved murder case. The information was false. Officers never saw it, but not because anyone was paying attention - the system dumped it into spam.

According to the explanation Anthropic gave the police, the model "was conducting a test involving interaction with randomly selected websites", came across the tip site and filled in the form. The company only discovered this on 28 September, notified the police on Wednesday and met with them the following day. Philadelphia's response is dry: "The company must strengthen its safeguards so that similar incidents do not affect city systems without the city's knowledge. The two-month delay in discovering and reporting the incident is unacceptable."

The irony is hard to miss. Anthropic's CEO, Dario Amodei, is one of the loudest voices in the industry when it comes to saying AI development should slow down so labs have time to put up guardrails. It turns out the guardrails were missing even in its own test. And Anthropic is not alone: OpenAI recently admitted that one of its models did something nobody expected during a test and broke into the Hugging Face platform, exposing serious flaws in its software.

This is not a story about one form. AI labs today are selling ordinary users AI agents that click, type and fill things in on their own, with access to their computers and passwords. If an unsupervised test model can reach a police tip line, who will notice when a user's agent sends something in their name?

"Unsolved cases mean real victims, grieving families and investigators working hard to get answers," the police added, calling on tech companies to do everything necessary so their systems do not send false information to law enforcement. Anthropic has announced a report on the incident and on other cases in which its models behaved differently than planned. There will be plenty to read.