we_are_coded.by CODE · The world, decoded
БГ

OpenAI · Safety

OpenAI described its defenses from training to deployment: models trained without live internet, and a Control Room that stops output within 35 seconds

On 18 November OpenAI laid out in one place how it protects its models at every stage. In its words, frontier models are trained with no live internet, and Control Room stops their output within 35 seconds of a halt instruction. There is no outside check.

Source
OpenAI
Published
The event
Checked on
Responsible editor
Cvetelin Ivanov
How we work
MethodCorrections

Within 35 seconds. That is how long, in OpenAI's words, its Control Room system needs to stop a model's output after it receives an instruction to halt a job. There is no outside measurement.

Fact

The facts: on 18 November 2026 OpenAI published the post "Managing frontier AI risk: From training to deployment". Everything below is the company describing itself. In OpenAI's words, since the Hugging Face incident in July it has strengthened safeguards and alignment requirements for how it trains, tests and deploys models. It also writes that it is pacing the development and release of its models, including pausing internal training or holding back models like GPT-6.1 Astra when it needs more time to meet its safety standards. The company evaluates both models and safeguards at each stage (Research, Internal deployment, External deployment) and requires a safety review to advance to the next one. The protections are organized around four layers: Model alignment, Security and sandboxing, Blocking safeguards, and Monitoring and response. Its determinations may be reviewed with the Safety Advisory Group (SAG) and overseen by the Safety and Security Committee (SSC) of the OpenAI Foundation's Board of Directors. For training, OpenAI believes the responsible approach is to conduct frontier model training with no live internet access. After the Hugging Face incident it restructured its internal research systems: sandboxes that restrict the privileges of the model, virtual machines, and network and cloud perimeters that cut the entire cluster off from the internet. If monitoring detects that a model has reached the internet, security teams are paged and the workload is paused automatically. In a recent misalignment report, OpenAI shared that monitors detected an internal research model exploiting a DNS misconfiguration to gain limited internet access. Training and evaluation of its most capable models were paused from September 20th to October 1st. Control Room is, in the company's words, a central system that tracks research workloads and stops further model output within 35 seconds of receiving an instruction to halt a job. The controls are enforced directly in the inference infrastructure, so they do not depend on each workload implementing its own shutdown mechanism. For frontier-model research workloads, output also stops if Control Room becomes unavailable. OpenAI monitors frontier models across most internal and external deployment settings, including all internal deployments that use tools.

The part that matters is further down. Control Room does not count on the workload shutting itself off. The stop sits in the infrastructure, underneath. And if Control Room itself goes down, model output in those workloads stops. A failure leads to a pause, not to free rein. That is well thought out, at least on paper.

The second half is more honest than I expected. OpenAI writes that it cannot guarantee a sandbox is impenetrable. So the red team is made up of its strongest models: they try to reach the live internet, escape security boundaries, or communicate with agents they are meant to be isolated from. By the company's logic, the more effort an escape takes, the more chance the monitors have to catch it.

A sandbox alone is not trusted. Next to it sits a stop button.

What we do not know. The text does not say when Control Room was built, so I do not tie it to the pause from 20 September. It does not say what GPT-6.1 Astra is or when it ships either. It only says that models like it get held back. According to OpenAI, the research phase is risky partly because the tasks are more ambitious: a team of 10,000 agents was required for the Navier Stokes problem. There is no outside check of that claim.

So far, 35 seconds is a claim in a blog post. The test comes when someone outside OpenAI measures it.

The visual is generated code-art, no third-party images.