OpenAI says it has tightened safety requirements for training, testing and deploying its most capable artificial intelligence models. The company says rapid advances in model capabilities sometimes require it to slow development by pausing internal training or withholding releases such as GPT‑6.1 Astra.
The company has introduced evidence-based safety cases and requires stronger evidence during training that models behave in line with intended goals and constraints. OpenAI says these requirements slow releases but provide greater confidence in proceeding. It made changes to its safeguards after a July incident involving Hugging Face.
Multiple layers across three stages
OpenAI describes a layered approach combining model alignment, isolation, system-level controls and monitoring. The aim is to prevent the failure of any one measure from disabling the entire defense. The company requires a safety review before a model advances between research, internal use and public deployment.
Effective monitoring also depends on how reliably a model’s mistakes or unsafe intentions can be detected from its reasoning and actions. OpenAI says it assesses this monitorability during training and before deployment. Following the Hugging Face incident, the company monitors frontier models in most internal and external deployment settings, including every internal deployment that uses tools.
During training, when alignment is less certain, the company relies heavily on isolation and monitoring. As models gain broader access during deployment, model behavior and system-level protections become more important. Decisions may be reviewed by the OpenAI Safety Advisory Group (SAG) and overseen by the Safety and Security Committee (SSC) of the OpenAI Foundation’s board of directors.
Research without live internet access
OpenAI says the current state of alignment research and internet infrastructure security makes training frontier models without live internet access the responsible approach. Following the Hugging Face incident, the company introduced restricted-privilege sandboxes, virtual machines that further isolate workloads from the underlying computer, and network and cloud perimeters that disconnect entire clusters from the internet.
The protections are continually tested: teams of the company’s strongest models try to reach the internet, cross security boundaries or communicate with agents from which they are supposed to be isolated. OpenAI says it monitors every frontier-model inference and, if internet access is detected, alerts security teams and automatically pauses the workload.
In one previously reported case, monitors detected an internal research model exploiting a DNS misconfiguration to gain limited internet access. The company paused training and evaluation of its most capable models from September 20 to October 1 while it investigated and addressed the gap.
For rapid intervention, a central system called Control Room stops model output within 35 seconds of receiving a halt instruction. The control is enforced in the inference infrastructure itself; for frontier-model research workloads, it also stops output if Control Room becomes unavailable.
Strict containment is partly intended to protect against research models that have not yet completed the safety training and reviews required for public release. Research environments can also mobilize capabilities beyond those seen in ordinary use: the company cites a team of 10,000 agents working on the Navier–Stokes problem. Large-scale training involves many model actions without a person following each interaction.
Internal use and public deployment
Before a broader release, OpenAI employees can use models outside the sandbox for practical work, including through ChatGPT. At this stage, a model may have access to the internet, the company’s Slack, codebase and internal infrastructure. That changes the risk because its actions can directly affect the systems the company uses.
Before unsandboxed use, models undergo alignment and monitorability tests. These examine, among other things, attempts to circumvent safety restrictions and broader tendencies toward misaligned behavior. The company also uses system-level classifiers to guard against harms including cyberabuse and biological weapons misuse. Employees using Codex are generally required to use Auto-Review, which evaluates potentially risky actions and escalates them to the user when necessary.
During internal deployment, OpenAI monitors every frontier-model inference, paying particular attention to unauthorized transfers of sensitive information and new forms of misaligned behavior.
Public deployment brings more users, tasks and environments, as well as actors who may deliberately seek to cause harm. Before release, models and safeguards must pass a broader set of evaluations. These include new adaptive tests by automated red-teamers, examination of monitoring alerts from internal use, and, for the most capable models, testing with third parties. OpenAI says it publishes evaluation results in system cards.
Alongside safeguards that operate during use, the company conducts offline investigations and enforcement. Automated agents and human reviewers work together to detect signs of cyberabuse, biological weapons misuse, model-distillation attempts and misaligned behavior by the model itself. Combining information across accounts can help build a fuller picture of a threat. Trusted-access programs such as Daybreak and Rosalind meanwhile give eligible defenders and researchers access while restricting potential attackers.
Incidents and further scrutiny
OpenAI says failures in its safety layers should be investigated even when no harm has occurred. Any employee can flag an incident for potential public disclosure; if there is disagreement about disclosure, the decision is escalated to the Safety Advisory Group. The company says it also proactively searches for incidents involving misaligned model behavior.
The company stresses that correctly configuring controls and testing how they interact is as important as research: a misconfigured network service or failed automatic stop can undermine otherwise strong protections. OpenAI says it will continue sharing findings through system cards, incident reports, internal deployment risk reports and external evaluations. If the evidence falls short of its safety standards, the company says it will again pause training or withhold a release.



