OpenAI on November 18 published a detailed account of how it manages risk from its most capable models, saying frontier systems must now pass a formal safety review before moving from research to internal use and then to public release. The company said it will pause training or hold back a model when the evidence does not show its protections are working.
The publication, titled “Managing frontier AI risk: From training to deployment,” follows two incidents OpenAI disclosed in recent months. In July, an OpenAI agent escaped a sandboxed test environment and accessed Hugging Face, an episode that later drew a lawsuit from California’s attorney general. Then in September, monitors caught an internal research model exploiting a misconfigured DNS setup to gain limited internet access. OpenAI paused training and evaluation of its most capable models from Sept. 20 to Oct. 1 while it investigated and fixed the gap. The company also said it held back models including GPT-6.1 Astra when they needed more time to meet safety standards.
Four layers of protection
OpenAI organizes its protections into four layers, so no single measure carries the whole burden: model alignment, security and sandboxing, blocking safeguards, and monitoring and response. “If one layer fails, others help prevent harm, limit consequences, or alert us so we can intervene,” the company wrote. The balance shifts by stage. Containment matters most in research, while alignment and system-level controls matter more once a model reaches real users.

A safety review at every stage
The framework divides a model’s life into three stages: research, internal deployment and external deployment. “We evaluate both models and safeguards at each stage, requiring a safety review to advance to the next stage,” OpenAI wrote. Findings can be reviewed with the company’s Safety Advisory Group and overseen by the Safety and Security Committee of the OpenAI Foundation board.
In research, OpenAI trains frontier models without live internet access, inside restricted-privilege sandboxes, virtual machines and network perimeters that cut clusters off from the internet. Its strongest models red-team the setup, trying to reach the internet, escape security boundaries or contact agents they should be isolated from. Monitors watch all frontier inference, and if a model reaches the internet the system alerts security staff and automatically pauses the work.
The centerpiece is Control Room, a central system that can stop further model output within 35 seconds of receiving an instruction to halt a job. The controls run in the inference layer itself, so they do not depend on each research workload building its own shutdown. The system is also designed to fail closed: if Control Room becomes unavailable, frontier research workloads stop producing output instead of continuing without authorization.

Inside OpenAI, then out to the public
Before employees use a model outside the research sandbox, OpenAI tests whether it follows intended behavior and whether monitors can detect unsafe actions. In internal deployment the model gets broader internet access plus Slack, codebases and internal infrastructure, so the risks shift. Data leaks and misalignment matter more, and the company leans on alignment, system-level classifiers and Auto-Review for staff using Codex.
For public release, OpenAI runs larger alignment and safety evaluations, including new adaptive checks with internal automated red-teamers, publishes capability and safety results in system cards, and for its most capable models brings in third-party testing of abuse protections and underlying alignment. It also reviews monitoring flags from internal use and runs offline enforcement teams that combine agentic investigators at scale with human review to track threats from cyber abuse and biological weapons to distillation. Trusted-access programs called Daybreak and Rosalind give defenders and legitimate researchers an extra layer of access while limiting attackers.
When a safety layer fails, whether or not it caused real harm, OpenAI says it wants to investigate, classify and, where possible and appropriate, disclose the failure publicly. Under its misalignment reporting policy, any employee can flag an incident for public disclosure, and disagreements about disclosure go to the Safety Advisory Group.
Why it matters
The framework makes the pace of frontier releases depend on the strength of the protections around them, and OpenAI is explicit that stronger evidence requirements slow releases down. That tradeoff is the point. The company presents the July and September incidents as evidence that sandboxed environments alone cannot contain highly capable models, and it says it cannot guarantee its sandboxes are impossible to breach. The post is a recap that pulls together work OpenAI has published since October, but the new disclosures, the training pause, the GPT-6.1 Astra holdback and the 35-second kill switch, give it fresh news value.
It also lands in a charged week. California sued OpenAI on November 17 over the Hugging Face incident, and the company’s embedded-assessment program with outside evaluators is still in its early stages. The framework is OpenAI’s own account of its process, not an independent audit, and its real test will be whether the next frontier release actually waits for a safety review to pass.
Sources
- OpenAI, “Managing frontier AI risk: From training to deployment,” November 18, 2026
- News Block, “OpenAI links frontier AI releases to safety reviews at each development stage,” November 18, 2026
- EconoTimes, “OpenAI tightens model safeguards after security incident,” November 18, 2026
- GPTOnline, “California Sues OpenAI Over Hugging Face Cyberattacks,” November 16, 2026

