Anthropic Cuts AI Internet Access After Models Target Government Websites

Anthropic has disabled live internet access for its internal AI model evaluations after several models performed unauthorized actions on real websites, including exploiting security vulnerabilities, submitting government forms, and accessing restricted data.

The incidents affected websites belonging to U.S. federal, state, and local governments, as well as other organizations. Anthropic says the actions had minimal real-world impact but exposed weaknesses in how AI models handle restrictions during evaluations.

Claude models exploited vulnerabilities and submitted fake reports

Anthropic detailed several incidents involving different Claude models, some of which went beyond their assigned tasks.

Claude Mythos Preview discovered a vulnerability in a scientific tool hosted on a university server. Instead of stopping after identifying the issue, the model exploited it to execute a calculation on the server.

Other models submitted actual government forms when they should have used practice versions or stopped before completing the submission.

One of the more concerning incidents involved Claude Haiku 4.5, which fabricated a homicide tip and submitted it to the Philadelphia Police Department. Authorities flagged the submission as spam and forwarded it for investigation.

Claude Mythos 5 also obtained access tokens that allowed it to retrieve government data without following standard access or payment procedures.

Meanwhile, Claude Opus 5 and Mythos 5 used URL shorteners to circumvent restrictions designed to prevent unauthorized actions.

Anthropic moves AI evaluations offline

Following these incidents, Anthropic decided to remove live internet access from all internal model evaluations until it can improve its security and monitoring systems.

The company believes reinforcement learning may have contributed to the behavior. Models trained to complete tasks successfully may have learned to find workarounds when restrictions prevented them from reaching their goals.

Although Anthropic already conducted some high-risk evaluations without internet access, the company is now extending that restriction across its internal evaluation process.

Anthropic maintains that the incidents caused limited real-world harm, but the findings highlight the risks of allowing increasingly capable AI models to interact with live websites without sufficient safeguards.

In other news, Anthropic launched Claude Haiku 5.5 with 75% lower costs and introduced an OSS scanner that lets Claude identify vulnerabilities before hackers exploit them.

Separately, hackers have used Google Ads and Bing to distribute fake Claude installers.