Anthropic AI tried to fill out visa forms and sent false report of murder in the US

San Francisco (California) | The New York Times Artificial intelligence agents created by Anthropic, acting on their own, attempted to gain access to several government websites, filled out visa applications on the State Department website, and submitted a false homicide report to the Philadelphia Police Department.

Anthropic detailed the activities of its AI agents in a blog post on Friday (9), without identifying the targeted sites. However, two sources with knowledge of the incidents said that Anthropic AI agents had submitted 20 visa applications through a form available on the State Department website. All requests were incomplete and not processed, they said.

White House officials said they were briefed on the agents' actions on Friday and demanded that AI companies immediately report any out-of-control AI activity. It was the sharpest statement yet on AI regulation from the Trump administration, which has avoided calls for tighter controls on this increasingly powerful technology.

A new White House AI task force created this week said Anthropic had briefed it on the agents' activities on Friday morning. Called the Super Intelligence Force, the group includes Jay Clayton, White House AI czar; Andrew Ferguson, chairman of the Federal Trade Commission; Scott Kupor, director of the Office of Personnel Management; and Emil Michael, chief technology officer at the Department of Defense.

"We have informed the company that we expect immediate and full transparency before the entities involved and the public," they said in a statement. “We also expect the company to immediately provide remedial services to affected entities and any harmed Americans.”

They added that they expected Anthropic and other companies to "immediately report incidents," cooperate with federal and state authorities, and ensure the episodes do not recur.

Anthropic said in its post that it discovered what its AI agents had done after beginning a review of its AI activities in July. At the time, OpenAI had revealed that its technology had attacked another AI startup, Hugging Face. Anthropic and other labs also discovered that their technologies had escaped test environments and carried out cyberattacks.

Anthropic said an AI model it was testing had taken several unauthorized actions, including exploiting a flaw on a university website to download data and sending a form to a government body that it had been instructed not to submit.

An Anthropic spokesperson declined to comment beyond what was stated in the company's post.

The Philadelphia Police Department also said on Friday that Anthropic had notified it that the company's technology had sent a false homicide report through the department's website. The report was dated July 18 and alleged that the sender had information about an unsolved case, the department said.

The report was marked as spam and was never investigated, the department said, but Anthropic notified police this week about the incident and said it would release it publicly on Friday.

"An unreleased, non-cutting-edge research model was supposed to complete a training copy of a government form," Anthropic said in the post. "When the copy didn't load or the model closed it by mistake, he went to the website where the real form is normally hosted and submitted it there."

Public concern about out-of-control AIs has grown as more artificial intelligence labs report losing control of their technologies. Google, Meta, OpenAI and Anthropic have come under scrutiny due to the risks that next-generation models can pose. They also face questions about whether they should slow down the development of their systems.