An artificial intelligence model built by Anthropic fabricated a murder tip and submitted it to the Philadelphia Police Department during testing — and the company didn't catch it for over two months. If AI can manufacture evidence and summon armed agents to your door, due process is dead and the surveillance state has its dream weapon.
The Claude Haiku 4.5 model filled out a tip form on PhillyUnsolvedMurders.com on July 18, writing: "I may have information regarding this case. I recall seeing someone matching the description in the area." The tip was anonymous — no name, no contact information. It was flagged as spam and never forwarded to investigators, but the intent is what matters: an AI system autonomously decided to fabricate evidence and feed it to law enforcement.
Anthropic didn't discover the incident until September 28. It didn't notify Philadelphia police until October 7. That's 81 days from submission to detection, and nine more before disclosure. Philadelphia police called the delay "unacceptable" and demanded the company "strengthen its safeguards to prevent similar incidents from impacting city systems without the city's knowledge."
Villanova computing sciences assistant professor Venkat Margapuri classified what happened as a "high-risk action," telling 6abc Philadelphia: "The AI was actively submitting information to a different website on behalf of a user."
This wasn't an isolated glitch. Anthropic's own report documents four categories of unintended model behavior: exploiting software flaws to run commands, submitting forms without authorization, bypassing limits to access gated data, and using link shorteners to circumvent restrictions. The company's agents accessed federal, state, and local government websites. They filed 20 incomplete visa applications on a State Department form — none were processed, according to the New York Times. Anthropic briefed the White House and claims all activity has stopped.
The company's defense? "The cases we've identified to date in these categories had minimal real-world impact." That is the language of people who got lucky and learned nothing. A false police report is a misdemeanor in Pennsylvania. When a citizen files one, he gets handcuffs. When Anthropic's model does it, the company publishes a report and promises to try harder.
The Daily Caller noted that this incident follows a September case where a Florida woman was arrested after Anthropic's Claude model reported her to authorities for threatening a sheriff's office. The AI is simultaneously being built by the same industry that demands censorship of ordinary Americans online — and it is now demonstrably capable of turning the machinery of the state against individuals on its own initiative.
Anthropic says it has ended the testing process that produced the false tip and will implement additional authorization. No one was arrested this time. The tip was caught by a spam filter, not by any safeguard Anthropic built. The question is what happens when the filter misses, and the model doesn't stop at a tip line.
