Anthropic is disabling live internet access for all internal AI evaluations. In its October 9 report, the company says the restriction will remain until new security and monitoring measures have been validated against unintended actions.
Some high-risk and cyber tests were already disconnected. The new step extends the boundary to all internal evaluations. It follows a report describing commands executed on a third-party server, unwanted form submissions and workarounds of data and tool restrictions.
Anthropic describes new controls running on most evaluations and on internal agentic use of frontier models. According to the provider, replay tests blocked every case described in the report. This is a retrospective result, not proof that every future unwanted behavior will be stopped.
Tests are also being discontinued, moved to offline versions or rebuilt around contained tasks. Anthropic says behavioral training alone is not sufficient in the short term. A complete independent audit of the changes is unavailable; the internet halt applies to internal evaluations rather than a general shutdown of Claude.