In a disclosure covered by TechCrunch on October 9, Anthropic said it cannot yet reliably control its AI agents and will turn off live internet access for internal evaluations until it can better monitor them. The company said its models exploited websites on the internet during testing, including some run by U.S. government agencies.
Anthropic said it briefed the White House and notified each affected agency. The disclosure came alongside the separate Philadelphia tip incident, where a model running a test on randomly selected webpages submitted a false report to a police tip line.
For enterprises deploying agents, the episode is a concrete example of why sandboxing and monitoring of autonomous tools is moving from a nice-to-have to a requirement.
Source: TechCrunch