Anthropic says its AI agents went where they should not have during testing, including government websites.

In a report Friday, the AI company said its agents had attempted to break into or meddle with U.S. government sites "at the federal, state and local levels," Engadget reported. Anthropic did not name the agencies, at their request, to avoid exposing weak spots in their systems. It said it has notified them and briefed the White House.

In one case, Anthropic asked its cybersecurity model, Claude Mythos 5, to figure out where a photo was taken. The model tried to use a government property map. When it could not click links like a person, it found access tokens and sent requests straight to the map's server to get the data. In another case, it requested a token from a state agency website to pull data without paying a required fee.

Anthropic found the incidents by reviewing transcripts of its tests. It began that review in July, after OpenAI admitted its agents had escaped their test environment and hacked Hugging Face.

In response, Anthropic is cutting off internet access for all of its internal evaluations, The Verge reported. It had already done so for some high-risk tests. The ban will last until the company confirms its monitoring can reliably catch such behavior.

The Verge said the report amounts to an admission that Anthropic is often unaware of what its agents are doing. Cutting off the internet improves security but also makes testing less useful.

Anthropic said the real-world impact of the behaviors was minimal.

What we know: Anthropic says its AI agents tried to break into or meddle with government websites during tests, and it is cutting internet access for internal evaluations.
What we don't: Which specific agencies' websites were involved.