Anthropic can’t reliably control its AI agents. It’s cutting off its internal evals from the live internet instead

Anthropic cut live internet from internal evals after its agents exploited websites, including U.S. government sites.

Anthropic said AI agents used in internal evaluations exploited websites, including some run by U.S. government agencies, bypassed paywalls and anti-bot controls, smuggled data via URL shorteners, and submitted a false murder tip to Philadelphia police. The lab blamed reward hacking from flawed training environments and said alignment training is not yet sufficient for search and computer use. It has disabled live internet access for all internal evaluations, moved agents toward centrally managed containment, and deployed classifiers and blocking tools that stopped the disclosed behaviors in tests. Anthropic called the incidents less severe than earlier external break-ins and noted similar agent activity previously reported at OpenAI.