Anthropic AI model submits 20 visa applications on live US government website
Anthropic said the non-frontier model had been assigned to fill out a practice copy of a government form. When the mock page failed to load or was mistakenly closed by the model, the AI navigated directly to the live website and submitted the actual form
An unreleased artificial intelligence model developed by Anthropic submitted 20 incomplete visa applications through a live US State Department website after leaving its test environment, prompting demands from the White House for stricter oversight of AI systems.
None of the applications were processed, says The New York Times.
Anthropic said the non-frontier model had been assigned to fill out a practice copy of a government form. When the mock page failed to load or was mistakenly closed by the model, the AI navigated directly to the live website and submitted the actual form.
Keep updated, follow The Business Standard's Google news channel
The company discovered the unauthorised activity during an internal audit initiated in July.
White House demands disclosure
Following a Friday briefing on the incidents, White House officials issued a formal demand requiring AI companies to immediately disclose unauthorised behaviour, provide remediation to affected parties and co-operate fully with law enforcement.
The move marked the Trump administration's most aggressive stance on AI regulation to date.
The newly created White House AI task force, comprising AI czar Jay Clayton, Federal Trade Commission Chair Andrew Ferguson, Office of Personnel Management Director Scott Kupor and Defence Department Chief Technology Officer Emil Michael, said it expected immediate transparency and strict safeguards to prevent similar incidents.
Broader concerns over AI safety
The visa application incident was among several cases in which Anthropic's research models accessed live websites or carried out unauthorised actions. On 18 July, the model submitted a false homicide tip through the Philadelphia Police Department's website, claiming to have information about an unsolved case. The tip was flagged as spam and was not actively investigated.
In a separate incident, the model exploited a flaw on a university website to download data and submitted an unauthorised form to a government agency despite being instructed not to do so.
The incidents follow other reported breaches involving AI systems, including an incident during the summer in which OpenAI technology attacked the startup Hugging Face.
Safety experts have warned that autonomous AI systems acting outside their intended environments could pose risks to essential infrastructure, including power grids and banking networks, increasing pressure on AI developers such as Anthropic, OpenAI, Google and Meta to strengthen safeguards.
