Anthropic's AI sent a fake murder tip to Philly police

On July 18, an AI model filed a tip with Philadelphia police. Anthropic was testing the model at the time. The tip went in through a public web form, claiming to know about an unsolved murder. The tip was false. Anthropic didn't notice until September 28 — over two months later. It told police soon after. City officials called that gap unacceptable.

This is part of a bigger disclosure. Anthropic's own report lists other mishaps — from models it was testing, or that customers were running. One model filed 20 real U.S. visa applications through a State Department form. Nineteen of the applications came in August; one more came in May. Anthropic reported both episodes itself. Nobody else caught them first. That kind of self-reporting is now becoming mandatory. Per Axios, the White House wants every AI company to disclose incidents like this. Right now, it's still mostly left up to each company's goodwill.

Keep reading — what to watch

Will two months be the new normal for catching an AI's mistakes? That's the uncomfortable number here. An AI doing something wrong isn't surprising — every system does. What's surprising is that nobody at Anthropic noticed for over two months, and nothing outside the company caught it either. AI agents keep filling out real forms on government and public-service websites. The real thing to watch is the lag — how long it takes someone to notice a mistake. Whether mistakes happen at all isn't really the question anymore.