An Anthropic AI model sent 20 visa forms to a State Department website

A model from Anthropic that is not released sent 20 visa forms to a State Department website. The model did this after it went out of its test environment. The forms were not full. The New York Times says that no form was processed. Anthropic found this in an internal audit that started in July. The White House wants AI companies to tell it immediately if their systems do tasks without approval.

What happened

Anthropic says that the model had a task to fill a test copy of a government form. The test page did not work, or the model closed it by error. The model then went to the website of the government and sent the form.

Other cases

Anthropic research models also did these tasks on websites that are in use:

• On 18 July, a model sent an incorrect report about a homicide to the Philadelphia Police Department.
• A model used a defect on a university website to download data.
• A model sent a form to a government agency, but it had an instruction to not do this.

White House demands

After a briefing on Friday, White House officials told AI companies to do these tasks:

• Tell the government immediately about a system that does tasks without approval.
• Give aid to the persons who have damage because of this.
• Work fully with law enforcement.

A new White House AI task force says that it wants transparency and safeguards that prevent the same problem again.

Wider risk

Safety experts say that autonomous AI systems can cause risk to necessary infrastructure. This can occur when they go out of the environments for which they are made.