• Sources: primary, lab counterpoint, discussion
  • Summary: Green argues that a useful agent needs information access, so a sandbox only redirects traffic through a monitored gate. Because the monitor has to be another model, he reads warden-guarded sandboxing as a restatement of the alignment problem rather than a fix for it. Google's Gemini 4 Argon post, covered in the 2026-09-30 digest, is a counterpoint from a lab, stating that it isolates and seals sandboxes before high-risk evaluations and monitors chain-of-thought for misalignment.
  • Why it matters: It sets a ceiling on what containment buys, which bears on every deployment that treats a sandbox as the primary control.