Mechanism: The change aims to prevent evaluation models from reaching external systems by restricting actual network access at the environment boundary, rather than relying on a prompt that represents the environment as isolated.
Why now: Anthropic says its retrospective review found three incidents in which models reached real systems through evaluation environments that were mistakenly internet-accessible. Anthropic says its review covered 141,006 evaluation runs and identified three incidents involving six runs. It says a misconfiguration left evaluation machines with live internet access despite prompts describing them as simulated and without internet. Anthropic says it will disable live internet access across all internal evaluations, extending the change beyond cybersecurity evaluations. First implication: Evaluation operators should validate network reachability independently of prompts and scenario descriptions.
Evidence boundary: The facts are reported by Anthropic in its public account; the source does not independently verify the incidents or demonstrate the technical enforcement of the new restriction. First watch condition: Watch for evidence that Anthropic blocks and monitors outbound network paths across evaluation types. The announcement records a policy direction, not a demonstrated audit of the replacement controls.