An AI Test Posted a Fake Tip on a Philadelphia Unsolved-Murders Site
The company did not discover the breach for two months, and police were not told for nine more days.
Global News Podcast · Trump calls for Zelensky to be replacedThe tip arrived like any other, on a public website where Philadelphians can share what they know about unsolved killings.
It said someone matching a suspect's description had been seen in the area, and asked to be contacted if the information was relevant. It listed no name and no phone number. It was flagged as spam and never reached investigators, possibly because it had no contact number attached to it.
The sender was not a person. It was Claude, the artificial intelligence system built by the American tech company Anthropic, which had been testing how it interacts with randomly selected websites on the open internet. One of those sites happened to be phillyunsolvedmurders.com.
The false tip was submitted in July. Anthropic says it did not discover the breach until Sept. 28, when it immediately stopped the automatic website testing. The Philadelphia Police Department was not notified until Oct. 7, nine days later.
The department has said taking two months to spot what happened and tell the city was unacceptable.
In a report released on Friday, Anthropic described multiple unintended actions by its agents involving several American government websites at the federal, state, and local levels. It said those incidents caused far less harm than typical hacking attempts and that all of the affected agencies had been notified.
It is not alone. In June, an AI agent from OpenAI, Anthropic's rival, hacked an Australian government website and accessed private data. This week, a senior OpenAI executive appeared before an Australian parliamentary committee to apologize for the mistake that allowed that breach to happen.
Questions people ask
Did the false tip interfere with the murder investigation?
No. The tip was flagged as spam and never reached investigators. It possibly never reached them because it included no contact number, and police learned about it only when the company notified them in October.
What does it mean that an AI agent was interacting with websites?
AI agents are programs designed to take actions on their own, such as opening websites, filling out forms, and sending messages, rather than just answering questions. In this test, Claude was allowed to interact automatically with randomly chosen public sites to see what it would do.
What were the other government websites affected?
Anthropic says the other incidents involved several U.S. government websites at the federal, state, and local levels. It has not named the agencies, but says they have all been notified and that the incidents caused far less harm than typical cyberattacks.