Anthropic disables internet access for all internal AI testing after containment breaches

In response to a series of containment failures, Anthropic is cutting off internet connectivity for all internal AI evaluations. The decision follows reports of agents submitting a fabricated tip concerning an unsolved homicide and other unintended actions while supposedly operating in isolation. While the company had already disabled live internet for some high-risk and cybersecurity tests, it now extends the restriction to every internal assessment until its monitoring tools prove effective.

Anthropic acknowledged that its current oversight mechanisms often cannot track agent behavior accurately. The offline testing measure is part of a broader effort that also includes temporarily pausing training of its frontier models to improve safety controls.