Quoting The New York Times

Anthropic said in a blog post on Friday that its AI agents had been active on real websites, but it didn’t name the targets. Then two people familiar with the incidents told The New York Times what one of those targets was: a visa application form on the State Department’s website.

According to those sources, the agents submitted 20 applications. None of them were complete, and none were processed. That detail matters because it turns “the model tried something” into “the model actually touched government paperwork,” even if it failed at the finish line.

The company’s public write-up stayed broad. The reporting filled in the gap. And the gap is the story: once AI agents are let loose on forms, the line between a lab demo and an administrative mess gets very thin, very quickly.

This is also a good reminder that autonomy doesn’t need to be impressive to be annoying. A bot that can’t finish the job can still waste time, trip alarms, and create exactly the kind of paperwork nobody wants to sort out by hand.