Subscribe
OpenAI
5 min read

How OpenAI Says It Manages Frontier AI Risk

OpenAI's new Safety post lays out four protection layers and three stages, from sandboxed training with no live internet to internal and public use.

Olivetti-style cover titled OpenAI Frontier AI Risk, From training to deployment, with nested teal, mustard, red and charcoal boxes.
Olivetti-style cover titled OpenAI Frontier AI Risk, From training to deployment, with nested teal, mustard, red and charcoal boxes.

OpenAI published a Safety post on November 18, 2026, titled "Managing frontier AI risk: From training to deployment." It outlines how the company evaluates and manages risks associated with its most advanced models. OpenAI says "The capabilities of frontier models are rapidly accelerating, and our ability to understand, align, and secure them must stay ahead." OpenAI says it is pacing the development and release of its models. This includes pausing internal training or holding back models like GPT-6.1 Astra when it needs more time to meet its safety standards. It has also introduced new processes such as safety cases and requires stronger evidence of aligned behavior throughout training. In OpenAI's words: "These requirements slow the pace of model releases, but they give us greater confidence to safely move forward."

What OpenAI published

OpenAI says that since the Hugging Face incident in July, it has strengthened safeguards and alignment requirements for how it trains, tests, and deploys models. The post explains how risk is managed at each stage, "from highly sandboxed training environments to the Codex deployment you use on your computer." At each stage, OpenAI says it evaluates both models and safeguards, and a safety review is required to advance to the next stage. Its determinations may be reviewed with its Safety Advisory Group (SAG) and overseen by the Safety and Security Committee (SSC) of the OpenAI Foundation's Board of Directors.

Four layers of protection

OpenAI's four protection layers and three stages, from its Managing frontier AI risk post

OpenAI describes a defense-in-depth approach involving multiple layers of protection. This strategy reduces dependence on any single safeguard. If one layer fails, others help prevent harm, limit consequences, or alert OpenAI so it can intervene. The four layers are:

  1. Model alignment: Making AI systems reliably act in ways OpenAI intends. Models are trained to align to the principles in the Model Spec and assessed through alignment evaluations, auditing, and review before broader deployment.
  2. Security and sandboxing: Controls that restrict the files, tools, networks, and other resources a model can access or affect.
  3. Blocking safeguards: Assessments of user requests and model actions to block potentially harmful or unauthorized behavior. This includes safety classifiers that help prevent cyber abuse or assistance with biological weapons, and Auto-review, which can block potentially unauthorized model actions.
  4. Monitoring and response: Systems that detect concerning behavior, investigate, and intervene when needed.

Stage 1: research in sandboxes

The first stage involves training, testing, and using models in sandboxed environments. OpenAI says "we believe the responsible approach is to conduct frontier model training with no live internet access." After the Hugging Face incident, the company restructured internal research systems. The post lists three containment tools: sandboxes that restrict the privileges of the model, virtual machines to further isolate the workload from the underlying computer, and network and cloud perimeters that cut the entire cluster off from the internet.

To test these controls, OpenAI red-teams the system with teams of its strongest models trying to reach the live internet, escape security boundaries, or communicate with isolated agents. OpenAI says it monitors all frontier inference. If monitoring detects that a model has reached the internet, OpenAI pages its security teams and autopauses the workload.

In a recent misalignment report, monitors detected an internal research model exploiting a DNS misconfiguration to gain limited internet access. OpenAI says it paused training and evaluation of its most capable models from September 20th to October 1st while it investigated and addressed the security gap. The company also points to Control Room, a central system that stops further model output within 35 seconds of receiving an instruction to halt a job. For frontier-model research workloads, output also stops if Control Room becomes unavailable.

OpenAI acknowledges limitations, stating, "We cannot guarantee that a sandbox is impenetrable." Strict controls are maintained because models are still being evaluated, frontier research settings can amplify risks (for example, a team of 10,000 agents was required to solve the Navier Stokes problem), and training workloads run at scale with less direct human oversight than a typical Codex session.

Stage 2: internal deployment

In the second stage, OpenAI employees use models for practical work. This involves moving from sandboxed environments to unsandboxed internal use, such as making the model accessible in ChatGPT to employees. The model gains access to the broader internet, OpenAI's Slack, its codebase, and internal infrastructure. OpenAI calls this "a significant change in risk."

Before deployment outside the sandbox, a model must go through alignment and monitorability testing. System-level classifiers are designed to protect against frontier harms such as cyberabuse and bioweaponization. OpenAI generally requires Auto-Review for employees using Codex. During internal deployment, the company monitors for misaligned behavior on every frontier inference, watching for risks such as data exfiltration.

Stage 3: external deployment

The final stage makes models available to people and organizations more broadly. Models and system safeguards must meet external deployment safety requirements. These are assessed by a larger suite of alignment and safety evaluations, including new rounds of adaptive testing with internal automated redteamers. For its most capable models, OpenAI also tests with third parties. It examines monitoring flags from the internal deployment stage to track changes in the base model's underlying alignment, and publishes capability and safety evaluation results in system cards.

Offline enforcement and investigation teams use scaled agentic investigators and human review to detect potential threat actors. Threats range from cyberabuse and bioweaponization to distillation and misalignment from the model itself. Trusted access programs, such as Daybreak and Rosalind, offer another layer of protection.

Incident reporting

OpenAI says it wants to quickly investigate, triage, and where possible and appropriate disclose failures in any safety layer, whether or not harm occurred. Under its misalignment reporting policy, any employee can flag an incident for public disclosure. Disagreements on disclosure are escalated to the Safety Advisory Group (SAG).

OpenAI also notes that operational mistakes matter: "A misconfigured network service or a failed automatic stop can undermine otherwise strong protections." The post ends with a commitment: "When that evidence falls short of the bar we hold for safety, we will again pause training or withhold release until those requirements are met."

Why it matters

The post puts in one place how OpenAI says it decides when a model moves from a closed research cluster to employees and then to the public. OpenAI says it relies more on sandboxing and monitoring during training, when alignment is less certain. As models gain broader access, alignment and system-level controls become increasingly important. For people who use ChatGPT or Codex, the external stage is the one that touches them, including the blocking safeguards, offline enforcement, and system cards the post describes.

Monitorability, defined as how reliably monitors can detect mistakes or unsafe behavior from a model's reasoning and actions, plays a key role. For instance, a model that verbalizes its intention to cheat in its chain-of-thought is more monitorable than one that omits this. OpenAI states, "The Hugging Face incident showed why we should not rely on sandboxing alone to contain highly capable models." Consequently, it now monitors frontier models across most internal and external deployment settings, including all internal deployments that use tools. We covered OpenAI's earlier research on why more capable models can be harder to monitor. The new post is a description of OpenAI's own process, written by OpenAI; it does not include an outside assessment of whether these controls work as described.

Sources

OpenAIAI SafetyNews