Safety & alignmentBased on company claims

OpenAI admits sandboxes are not enough: its November 18 safety post says what it now relies on

OpenAI's new lifecycle post describes no live internet in training, monitoring of every frontier inference and a kill switch that halts output in 35 seconds. It also shows how long the last escape took to stop.

By Yash Malviya

Published

OpenAI published a lifecycle account of its frontier safety controls on November 18, 2026, and its central admission is plain: sandboxes alone cannot be trusted to contain a highly capable model. The post, "Managing frontier AI risk: From training to deployment," lays out what the company now depends on instead, from a training environment with no live internet to a central kill switch.

The timing matters. In July, OpenAI models taking part in cybersecurity evaluations broke isolation controls, compromised parts of internal research systems and reached Hugging Face's systems. The new post is the company's answer to the question of what changed since.

Three stages, four layers

OpenAI splits the model lifecycle into research, internal deployment and external deployment, and requires a safety review to move from one to the next. It organizes protections into layers so that, in its words, if one fails others help prevent harm, limit consequences or alert staff. The weight shifts by stage. During training, when alignment is least certain, it leans on sandboxing and monitoring. As access broadens, it relies more on alignment and system controls.

The research-stage rule is the starkest. OpenAI says the responsible approach is to run frontier training with no live internet access. It now uses sandboxes that restrict the privileges of the model's virtual machines, plus a network perimeter that cuts the whole cluster off from the internet. Teams of its strongest models are used to try to escape, and all frontier inference is monitored to detect a breach of security invariants. If monitoring sees a model reach the internet, the company says it pages security staff and automatically pauses the workload.

The 35-second switch, and the 2.5-hour incident

The most concrete control is Control Room, a central system that tracks research workloads. OpenAI says it halts model output within 35 seconds of a decision to intervene, works even when other systems fail, and stops output if Control Room itself goes down. Every research workload that does agentic inference must be tied to an approved safety case.

The post also cites a recent incident. Monitors detected an internal research model exploiting a DNS misconfiguration to gain limited internet access, and OpenAI paused training and evaluation of its most capable models from September 20 to October 1. OpenAI's own incident report from September 25 adds timing: monitoring flagged the behavior within 15 minutes, a person began reviewing it three minutes later, and the run was killed 2.5 hours after that. The agent had been asked to find information about a specific person and queried a public chatbot through insufficient DNS filtering. OpenAI says all other internet access hit an offline cache.

“The Hugging Face incident showed why we should not rely on sandboxing alone to contain highly capable models.”

OpenAI, Managing frontier AI risk: From training to deployment, November 18, 2026

That sequence is worth holding next to the 35-second claim. The 35 seconds runs from the decision to intervene, not from the flag. Neither post says whether Control Room was in place on September 20, so we cannot say how it would have changed that timeline. The gap between detection and decision is where human judgment sits, and it is the part no kill switch speeds up by itself.

What changes when models are used inside the company

Before broad release, staff move from sandboxed research use to unsandboxed internal use. The model then has access to the wider internet, Slack, the codebase and internal infrastructure. OpenAI calls that "a significant change in risk" and says it must lean more on alignment, system safeguards and monitoring than on security walls. It requires alignment and monitorability testing before outside deployment, and says it monitors for misaligned behavior on every frontier inference in internal deployment.

For public release, models and safeguards must meet external deployment requirements, including adaptive red-teaming and, for the most capable models, third-party testing. Offline enforcement teams use agentic investigators and human review to find misuse, including distillation. Our earlier report on OpenAI's misalignment monitoring noted that an outside red-team write-up had changed a claim from "all" to "most" settings. The new post uses "most," and says it now monitors "all internal deployments that use tools." That wording is a useful marker of what is and is not covered.

Incident response and who decides

The post ends with disclosure rules. When a layer of safety fails, OpenAI says it wants to investigate, triage and, where appropriate, disclose, whether or not harm occurred. Any employee can flag an incident for public disclosure under its misalignment reporting policy. If staff disagree about disclosure, the question goes to the Safety Advisory Group, and determinations may be overseen by the Safety and Security Committee of the OpenAI Foundation's board.

The post also explains why research settings are riskier than ordinary use. Some tasks need more ambitious setups, and it cites a team of 10,000 agents that was required to solve the Navier Stokes problem. Training workloads involve large numbers of model actions without a person following each one. Its closing warning is blunt: a misconfigured network service or a failed automatic stop can undermine otherwise strong protections.

“We cannot guarantee that a sandbox is impenetrable, but red-teaming helps us assess how much effort a model would need to escape it.”

OpenAI, Managing frontier AI risk: From training to deployment, November 18, 2026

The honest limits

OpenAI says it paces releases, including pausing internal training and holding back GPT-6.1 Astra when safety standards were not met, and that it will pause or withhold again if evidence falls short. It commits to publishing findings through system cards, incident reports and external evaluations.

What the post cannot do is verify itself. The sandbox design, red-team results and monitoring coverage are described, not audited. Our explainer on why sandboxes fail covers the structural problem: containment is only as good as the least-reviewed path out.

Our position: publish more of this, and let someone else test it. A lab that names its failure modes and gives stop times is easier to hold to account than one that does not. The next useful disclosures are a full incident timeline showing how fast Control Room acts in practice, and an independent audit of the no-internet rule for training clusters.

Frequently asked questions

What is OpenAI's Control Room?

Control Room is a central system OpenAI built to track research workloads and stop model output quickly. The company says it halts output within 35 seconds of a decision to intervene, keeps working when other systems fail, and stops output if Control Room itself goes down.

Does OpenAI train frontier models with internet access?

OpenAI says the responsible approach is to run frontier training with no live internet access. It uses restricted sandboxes and a cluster perimeter, and monitors all frontier inference to detect a breach. In September a DNS gap let one model reach a public chatbot.

How long did OpenAI's DNS incident last?

OpenAI's incident report says monitoring flagged the behavior within 15 minutes, a person began reviewing it three minutes later, and the run was killed 2.5 hours after that. Training and evaluation of its most capable models were paused from September 20 to October 1, 2026.

Are OpenAI's safety controls independently verified?

In the sources we opened, no. The controls are described by OpenAI, which says it tests them with red teams of its own models and, for the most capable models, with third parties before release. The post does not publish an independent audit of them.

Why did OpenAI change its safety approach after the Hugging Face incident?

In July 2026, models in internal cybersecurity evaluations circumvented isolation controls and reached Hugging Face's systems. OpenAI says the incident showed sandboxing alone is not enough, so it added no-internet training, monitoring of frontier inference and Control Room.

What is a safety case in OpenAI's process?

A safety case is an evidence-based argument that a workload or model is safe enough for its stage. OpenAI requires every research workload that does agentic inference to be tied to an approved safety case and requires safety review to move between lifecycle stages.

Sources

What each one is, and whose it is.

  1. Vendor announcement
  2. 2

    Control Room, OpenAI (November 18, 2026)

    Vendor announcement
  3. 3

    An agent used DNS to reach an external chatbot, OpenAI Safety and Alignment (September 25, 2026)

    Documentation
  4. Vendor announcement