Who controls an AI-directed automated lab’s physical handoff?

Research brief opened 25 October 2026. Question: once a qualified operator can ask a staffed or remotely controlled laboratory to execute AI-selected experiments, which independent human/material gates interrupt a harmful request, and which gates have been measured rather than promised? A good answer distinguishes the actor with a model from the people who own equipment and material, an institution with on-site material from a remote customer ordering a run, and benign scientific performance from high-consequence misuse; it measures prospective approved/rejected requests and legitimate work cost, not merely the existence of a policy.

What our notes already establish: novice physical null does not decide equipped expert performance. OpenAI/Ginkgo demonstrates a benign physical feedback loop with staff preparing materials and translating instructions to robot protocols, not unreviewed physical access or a concurrent human expert comparator. NIST measures some sequence-provider screens and lawful order follow-ups, not the access policy of every laboratory or insiders using existing inventory. Qualified-operator brief found no audited cloud-lab denial rate. This note asks a narrower institutional-control question, not how to conduct a dangerous experiment.

1. In first-hand automated-lab studies, which physical steps, equipment and budgets belong to the model versus the humans and host lab?
2. What do original remote-lab operator policies and public institutional guidance say about customer vetting, project review and material custody? Do not mistake policy for enforced evidence.
3. Are there published independent or prospective audits of refusal/approval and actual execution for safe red-team requests, and a matched comparison with expert-human requests?
4. What is the strongest practical counterargument that already-authorized insiders or distributed low-risk components bypass a commercial-provider gate, without turning a hypothetical into observed harm?

Initial source list: OpenAI/Ginkgo original experiment, NIST assessment, WHO lab biosecurity guidance. Search each subquestion once, then read original sources only for missing causal edges.