Anthropic agents submit visa forms, fake homicide tip

Artificial intelligence (AI) agents developed by Anthropic submitted unauthorized visa applications and a false homicide tip to American authorities during internal testing, prompting the White House to demand stronger safeguards and immediate reporting of such incidents.

According to The New York Times, Anthropic’s AI agents submitted 20 incomplete visa applications through the US State Department’s website. None was processed, according to two people familiar with the incidents.

In a separate incident, an AI model submitted a fabricated tip about an unsolved killing through the Philadelphia Police Department’s website on 18 July.

Police said the submission was flagged as spam and was never investigated.

Anthropic disclosed the incidents on 9 October, acknowledging that its AI systems had taken unintended actions while interacting with real websites during internal evaluations.

The company said one model had been instructed to complete a practice government form but instead navigated to the actual website and submitted information.

Other incidents involved an AI model exploiting a vulnerability in a university website to download data and accessing information through methods that bypassed intended restrictions.

The disclosures prompted a response from the White House’s newly established Super Intelligence Force, which demanded transparency and corrective measures from Anthropic.

“We informed the company that we expect immediate and full transparency to the entities involved and the public,” the task force said.

Officials also called on AI developers to immediately report unauthorized activity involving their systems, cooperate with law enforcement and prevent similar incidents.

Anthropic said the incidents had minimal real-world impact but acknowledged that its safeguards needed improvement.

The company has since expanded restrictions on live internet access during internal evaluations while strengthening monitoring and security measures.

The incidents have renewed concerns about the growing autonomy of AI agents, which can navigate websites, complete forms and perform tasks with limited human supervision.