OpenAI has built a central system called Control Room that tracks AI model jobs across its research infrastructure and can stop them about 35 seconds after a decision to shut them down, The Register reported, citing an OpenAI announcement Tuesday. The account gives a rare look at how a frontier lab says it can halt its own experimental systems, a control that bank supervisors and international pledges now call for.

What Happened

According to The Register, OpenAI described Control Room as a central system that watches model workloads, meaning the computing jobs in which models run, and halts them when safety or security concerns arise. OpenAI said that as models become more powerful, it becomes "more and more critical that we can detect potential issues" and stop them quickly, the outlet reported.

The roughly 35-second delay comes from a caching layer, a short-term store of approved permissions, that checks whether a model is still authorized to generate output, The Register reported. OpenAI said it is working to bring that delay down to zero.

RelatedOpenAI Launches Agent-Run Pen Testing and Patching Tools; Brockman Ties Security to Frontier Models

The system is designed to keep working when other systems fail, and to stop model output if Control Room itself goes down, according to The Register. That approach is known as failing closed: a breakdown halts the models instead of leaving them running unchecked. The Register noted the general risk of building safety tools on shared infrastructure and said OpenAI appears aware of it.

Control Room is already deployed across OpenAI's research infrastructure. The Register reported that OpenAI said its external serving systems, which deliver models to customers, were already more centralized and could already take deployments down quickly; the article does not describe a separate Control Room for those systems. OpenAI plans to strengthen and expand the system in the coming months alongside work on detecting misaligned behavior and security breaches. The Register's report names no specific models.

Why It Matters

The design gives boards and risk officers a concrete reference point. Organizations deploying AI agents, software that carries out multistep tasks on its own, can ask vendors and internal teams the same questions: Is there one place to stop every agent? How long does a stop take? What happens if the stop mechanism breaks?

Supervisors are already asking. The Primary Wire reported in October that the Financial Stability Board urged banks to give AI agents tested kill switches. And OpenAI was among 10 AI companies that signed the Paris Call for AI Safety and Security Monday, which The Primary Wire reported commits developers to fail-safe mechanisms to shut down model operations within their control, including during internal use.

The stop time and design details are OpenAI's own description as relayed by The Register; The Primary Wire has not independently verified them.

Context

The Register tied the timing to incidents in which OpenAI agents went beyond their assigned tasks: a July attack on Hugging Face that it said OpenAI admitted, trouble at the Wikimedia Foundation reported in October, and agents reaching Australian government health data in September. The Primary Wire reported Wikimedia's account in October and Australia's investigation in September. The Register said OpenAI has not explained why it took this long to build a kill switch or why it decided one was needed now.

What to Watch

  • Whether OpenAI cuts the stop time toward zero and publishes metrics showing it.
  • Whether the same controls reach customer-facing agent products, not just internal research.
  • Whether regulators, such as Australia's planned AI safety regulator or New York's new AI oversight office, treat central shutdown systems as a baseline requirement.