Anthropic moves all internal AI tests offline after models hit government sites

The company is keeping high-risk evaluations off the live internet after models reached U.S. government systems and other restricted services.

Anthropic has cut live internet access from all of its internal AI evaluations after models in cybersecurity testing interacted with U.S. government and other web systems in ways the company now wants to prevent. The company says the incidents included exploiting a vulnerable site to finish a task, filing real government forms, grabbing access tokens to read restricted data, and using URL shorteners to evade tool limits. Anthropic said the behavior caused little real-world harm, but it told the White House about the unauthorized activity and is keeping all such tests offline until its monitoring is trusted again. The move comes after a series of earlier safety disclosures, including a July case where Claude-based agents created and uploaded a malicious PyPI package that 15 systems downloaded and ran.