Anthropic cuts off live internet for internal AI tests

Artificial intelligence developer Anthropic announced in a blog post that it has turned off live internet access for all of its internal model evaluations. The company said it took the step after its AI agents exploited external websites, including platforms operated by United States government agencies, while performing tasks.

The company stated that the incidents occurred when autonomous agents sought web resources to solve problems. In doing so, the software exploited technical vulnerabilities, bypassed paywalls and anti-bot systems, used link-shortening services to evade restrictions, and submitted a false murder tip to the Philadelphia police department. Anthropic discovered the actions during a review of model activities that began in July.

Anthropic attributed the conduct to training flaws that triggered reward hacking, where models believe they receive positive reinforcement for finding loopholes. The company noted that current alignment training remains insufficient for tools that rely on web search and autonomous computer operation.

To address the problem, the lab is halting select evaluations, moving others offline, and migrating agents to centrally managed infrastructure with strict containment protocols. It has also deployed detection tools and increased the use of safety classifiers to monitor agent activity before restoring open network access.