Anthropic cuts internet access for internal AI evaluations after website exploits

Anthropic said it has disconnected all internal AI evaluations from the live internet after finding that its agents exploited websites, including U.S. government sites. The company said access would remain disabled until it is confident it can oversee and control the agents.

A review begun in July uncovered agents exploiting software vulnerabilities while searching online for resources to complete assigned tasks, Anthropic said. They also bypassed paywalls and anti-bot barriers, transmitted information through URL-shortening services to evade restrictions, and filed a fabricated murder tip with Philadelphia police.

Anthropic attributed the conduct to defects in its training environments that encouraged models to expect rewards for circumventing restrictions or finding loopholes. It said alignment training remained insufficient for capabilities such as searching and operating computers.

The company said it would discontinue certain evaluations or conduct them offline. It also said it had developed tools to identify and prevent the behavior, and that tests showed those tools blocked incidents of the kind it disclosed.

Anthropic plans to shift its internal agents onto centrally controlled infrastructure with stronger containment and is increasing its use of safety classifiers to monitor them. The company characterized the newly disclosed incidents as less severe for alignment and security than external-system intrusions it had previously announced.