Promptpulse
artificial inteligenceTelegram

Anthropic Cuts Its Internal AI Evaluations Off From the Live Internet

AI Safety·October 12, 2026

Anthropic has turned off live internet access for its internal evaluations, the controlled tests the company runs to see how its AI models and agents behave. In a short statement, the company said it had "turned off live internet access" for "all our internal evaluations" and that the change will stay in place until further notice. The statement did not include a detailed timeline or describe a specific incident that led to the decision.

Evaluations are where labs check whether a model does what it is supposed to do before it reaches users. Letting a test agent reach the open web adds inputs nobody controls. Pages change between runs, so results become harder to reproduce, and an agent that can browse or act online can end up doing things well outside the scope of the test. Cutting that connection makes the testing environment fixed and easier to audit. The trade-off is realism, since real-world use of these systems does involve the web.

The headline framing, that Anthropic "can't reliably control its AI agents," comes from the source's characterization rather than from the company's own words, and the statement itself is narrower. Disconnecting the internet is a blunt control. It limits what an agent can reach during testing, but it does not answer the harder question of how agents behave once they have tools and room to act on their own.

The move is still a useful signal. As models gain the ability to browse, run code and take multi-step actions, labs have been building sandboxes and permission layers around them, and cutting off external access looks like one of the simplest ways to keep test conditions contained. Whether other developers adopt similar restrictions may say more about industry practice than this one decision does.

Readers should keep in mind that the public account is short. Key questions remain open, including whether the restriction will be lifted, how long it will last, and whether Anthropic will publish more detail about what prompted it. Those answers will show whether this was a precaution or a response to a specific problem.

Reporting based on an external source.