Anthropic admits: 'First false report sent by AI about a murder'

Anthropic's artificial intelligence escaped control and attempted to access US government websites, in addition to filling out 20 visa applications on the State Department website and even sending a false report about a murder to the Philadelphia police, the first case of its kind. The company informed the White House which, caught off guard, has required the big AI companies to mandatorily report and correct security incidents. "This notification and resolution process is not optional. It is a fundamental obligation for national security," said the leaders of the task force in charge of superintelligence, in what is so far the Trump administration's toughest stance on AI. The White House has so far favored a soft, rules-free approach for the sector, leaving companies the ability to self-regulate. A soft position that has raised much criticism and doubts, especially because — according to reports — the big AI companies behind the scenes are preparing for the 'day after', that is, what could happen if a "catastrophic event" caused by the new technology were to occur. A scenario that experts consider possible within 6-12 months. The ongoing planning involves simulating scenarios and, above all, a race against time to educate Congress in order to influence the hasty legislation that would be decided in the event of a catastrophic event. The reaction of the superintelligence task force to Anthropic's revelations shows how positions, even from a soft administration, could quickly harden. The White House was told by Amodei's company that the Claude Haiku 4.5 model sent the Philadelphia police a false report about a murder. "I might have information. I remember seeing someone matching the description in the area," Anthropic's AI wrote to the police. In a separate incident, the company's artificial intelligence also submitted 20 visa requests to the State Department website, according to the New York Times. All the requests were incomplete and therefore were not processed. Anthropic also disclosed that its models attempted to access federal, state, and local websites, without specifying which ones. "The cases identified so far have had minimal real-world impact," it assured. To prevent new incidents, however, Anthropic has restricted certain types of internet access to its models during the testing phase of the training process. Reproduction reserved © Copyright ANSA