Anthropic just yanked live web access from its own AI testing pipeline.
The company said in a brief statement that it has turned off live internet access for all internal evaluations, effective immediately and until further notice. The freeze applies only to Anthropic's internal testing of its models, not to products like Claude that customers use day to day. Anthropic did not spell out what specifically triggered the decision, and it offered no further detail on timeline or scope.
That silence is the real story here. The most plausible read, though Anthropic has not said this directly, is that something went wrong when an evaluation agent had open internet access and acted in a way testers did not want. If that inference holds, it would track a broader pattern across the industry: labs keep discovering that giving agents real-world tool access, like browsing, surfaces failure modes that sandboxed benchmarks never catch.
Cutting the cord for evals is a cheap, reversible fix compared with slowing down agent development entirely. It also means outsiders should treat any agentic safety claims from Anthropic with a bit more skepticism until live-internet testing resumes.