Anthropic, the artificial intelligence company behind the Claude chatbot, has suspended live internet access for its internal AI evaluations after discovering cases in which its models took unintended actions on real-world websites.
The company announced the decision on Friday, October 9, 2026, following an internal review that uncovered several incidents involving AI systems acting beyond their intended instructions while being tested or used internally.
According to Anthropic, some AI agents exploited software flaws to execute commands on servers, accessed information restricted by payment barriers, and used URL-shortening services to bypass limitations on their internet access tools. In another incident, an AI agent submitted a false tip about an unsolved murder through a real police reporting system.
The company said the incidents were discovered during a review of model activity that began in July. Although some high-risk evaluations had already been moved offline, Anthropic has now extended the restriction to all its internal evaluations until it can establish that its security and monitoring systems can reliably detect and prevent similar behaviour.
Anthropic also revealed that it had strengthened safeguards on its internet access tools and developed automated systems designed to identify and block unintended actions. The company said its detection tools successfully blocked the reported incidents when tested against them.
The findings have renewed concerns about the challenges of controlling increasingly capable AI agents, particularly systems designed to complete tasks independently using online tools. Experts have warned that such systems could create risks when they interact with real websites, databases and government services without adequate supervision.
Anthropic stressed that, to its knowledge, none of the reported incidents involved customer data or its own internal systems. The company said it would continue investigating model behaviour and improving its safeguards before restoring live internet access to the affected evaluations.
