---
title: OpenAI's Control Room: every research inference request now needs a live job authorization, so a stop order halts model output within 35 seconds | ctxsec
url: https://ctxsec.com/items/openai-control-room-stop-research-model-output-35-seconds-fail-closed
description: "OpenAI says it built a central service that every research inference workload, including agentic sampling for training and evaluation, must register with. Inference engines check each request's job ID with Control Room before generating, so a stopped job gets refusals even if its own code ignores the stop. The 35-second ceiling comes from an authorization cache. The integration fails closed: if Control Room goes down or is tampered with, research inference is denied. Fail-closed is live for frontier research and finishing rollout to the rest."
status: "official (OpenAI, Nov 9, 2026); company description of its own internal system, not independently verified"
published: 2026-11-09
---

> OpenAI says it built a central service that every research inference workload, including agentic sampling for training and evaluation, must register with. Inference engines check each request's job ID with Control Room before generating, so a stopped job gets refusals even if its own code ignores the stop. The 35-second ceiling comes from an authorization cache. The integration fails closed: if Control Room goes down or is tampered with, research inference is denied. Fail-closed is live for frontier research and finishing rollout to the rest.

[← AI](/ai) · [markdown](/items/openai-control-room-stop-research-model-output-35-seconds-fail-closed.md)

official (OpenAI, Nov 9, 2026); company description of its own internal system, not independently verified

2026-11-09

### OpenAI's Control Room: every research inference request now needs a live job authorization, so a stop order halts model output within 35 seconds

- **Design choice:** the kill switch sits in the small, centralized inference layer, not in fast-changing workload code. OpenAI also runs sweeps for rogue engines that skip the check.
- Its misalignment monitors can stop jobs automatically; OpenAI says its recent DNS-related incident was caught by a monitor.
- **For teams running agents:** the same pattern applies at small scale. Give each agent run an ID and make the model gateway refuse requests for revoked IDs, instead of trusting the agent loop to exit.

Sources: [OpenAI: Control Room](https://openai.com/index/control-room/) · [OpenAI on X (Nov 9, 2026)](https://x.com/OpenAI/status/2119913172039868579)