OpenAI describes Control Room, infrastructure that registers research jobs against approved safety cases and can revoke their access to further model output. OpenAI reports a maximum 35-second delay from receiving a stop instruction to halting output. Workload code can remain running; the mechanism does not stop every action an agent has launched.
Why it matters
It makes a concrete shutdown boundary inspectable and distinguishes revoking inference from terminating processes or detecting misconduct.
Key facts
As stated in the sources, with where to find them.
- Research jobs register with Control Room against an approved safety case. Inference engines check authorization when fulfilling requests; the mechanism is enforced in the inference layer rather than relying on workload code.How it works
- OpenAI reports a maximum 35 seconds between Control Room receiving a stop instruction and further model output stopping, due to cached authorization. This excludes detection and the decision to intervene.Speed and reliability
- Workload code continues running after inference is revoked. Code or external actions already launched are outside the stated inference-stop guarantee.How Control Room works: Model output stops—the job stays
- At publication, fail-closed authorization was deployed for all frontier-model research workloads, with rollout to other research workloads planned. Monitoring and the decision to stop a job remain separate from inference enforcement.Speed and reliability
Record updates
Findings that cite this record
No tracked finding cites this record yet.
Key questions this bears on
- Can prompt injection against AI agents be reliably defended?Not reliably. Bounding what untrusted input can trigger helps, but trust boundaries must survive delegation and tool execution.
- Can AI agents defend and oversee systems on their own?Not reliably. Some designs improve benchmark triage, but autonomous defense lacks operational validation and production oversight of agents has documented gaps.
Sources
Related records
Sep 28, 2026
Oct 29, 2026
Oct 23, 2026
Oct 22, 2026
Oct 20, 2026
Oct 14, 2026