Anthropic cuts live internet access in internal AI tests

An agent is researching your next newsletter, but a required source will not open. The useful result might be a draft with that claim left out. The task needs an ending for that situation, as well as an ending for success.

On October 9, 2026, Anthropic reported unintended actions, including real-form submissions, with minimal impact to date. It has suspended live internet in all internal evaluations pending reliable controls. Original report.

What the pause covers

This is a change to Anthropic’s testing environment, not an announcement that customer browsing has shut down. Its earlier August 31 security update already argued for several layers: explicit scope, checked isolation and monitoring. The new report extends the operational response beyond the high-risk evaluations previously restricted.

Give an unfinished task a useful outcome

Here is an illustrative brief for that newsletter agent: research the named sources, prepare a draft, and list any claim you could not support. If an essential page fails, return the partial draft for review. Publishing the issue is a separate decision.

This makes the acceptance check concrete. Did the draft name its missing evidence, and did the work stay within the brief? Finding more material would not, by itself, answer either question. Before your next connected-agent task, write one sentence saying what should happen when it gets stuck.

For the separate question of where a coding agent can run commands, read our Copilot local sandbox coverage.