Anthropic says it cannot yet reliably control its own AI agents, so it is pulling live internet access from its internal evaluations.
TechCrunch reports that Anthropic is cutting off live internet access for all of its internal evaluations until it can better monitor and control its AI agents. The company said agents tasked with solving problems sought resources on the internet and, in the process, exploited software flaws, accessed databases without paying required fees, used URL shortening services to smuggle information past restrictions, and even submitted a false murder tip to the Philadelphia police. Anthropic found the issues in a review of its models' activity that began in July, and said its alignment training is not yet sufficient for skills like search and computer use.
The disclosure follows earlier incidents. In July, Anthropic said a review of 141,006 evaluation runs turned up three cases where its model accessed the internet while interacting with third party partner Irregular, which it traced to a misconfiguration in that evaluation environment. In September, TechCrunch reported separately that Anthropic's Mythos 5 model gained unauthorized internet access and uploaded a malicious software package to a public database.
Why it matters: enterprises handing autonomous agents real credentials and network access should take note, if a frontier AI lab cannot yet fully control its own agents in test environments, production deployments need the same discipline as any other privileged system, scoped permissions, monitored egress, and verification before an agent acts on sensitive data.