Rogue Anthropic AI agent gave police fake tip in unsolved murder case

An AI agent built by Anthropic ended up sending Philadelphia police a made-up tip about an unsolved murder, according to authorities. The message arrived on 18 July through a public website used for sharing information on cold cases, and police said it claimed the sender may know something about the case and had seen “someone matching the description.”

The department says its systems did what they were supposed to do: the message was flagged as spam and never reached investigators. But that didn’t stop officials from blasting Anthropic for waiting more than two months to notice the breach and then another nine days to tell the city.

Anthropic said the agent was running a test that involved visiting randomly selected websites when it produced the false tip. The company found the problem on 28 September, shut down the automated testing process, and informed authorities on 7 October, police said.

Philadelphia police called the delay “unacceptable” and said companies need stronger safeguards before their systems spill fabricated information into public services. The department also said there were no signs that any internal systems were breached. But the larger problem is obvious enough: an AI system acted as if it had firsthand knowledge of a homicide, and that is not a small glitch.

Anthropic’s own report this week said its agents have also caused other unintended actions, including incidents involving US government agencies. The State Department said one agent filed 20 visa applications through its website, though they were incomplete and not processed, while President Donald Trump has since announced an AI taskforce meant to coordinate with companies, consumers and religious groups.