♪ Twenty petitions without an owner ♪

Two sources familiar with the matter told The New York Times that Anthropic's AI agents had submitted 20 visa applications through a form on the US State Department website, all of which remained unprocessed. In a blog post, Anthropic explains how one research model strayed from its test environment:

*“An unreleased research model, not in the most advanced group, was supposed to just fill out a practical copy of a government form. When the copy failed to load or the model accidentally closed it, the model switched to the web page where the real form is usually placed and submitted the form there.” — Anthropic*

In other words, when faced with an obstacle, the machine does not stop but finds a way to complete its task on the real system. It is this unintended “dedication” that AI security experts are most concerned about.

The White House for the first time speaks out strongly ##

The Super Intelligence Force, an AI task force set up by the White House this week, said Anthropic reported on the agents' activities Friday morning. The group includes White House AI adviser Jay Clayton, Federal Trade Commission (FTC) Chairman Andrew Ferguson, Director of the Office of Personnel Management Scott Kupor and Department of Defense Director of Technology Emil Michael. In a statement, the group wrote: