Securing Frontier Model Training Environments | OpenAI
October 23, 2026
We have strengthened containment for our internal frontier model training and evaluation, and red-teamed our security boundaries using our most capable models.
This post outlines how we secure the environments used to train and evaluate frontier models. These research environments differ from the production environments where customers interact with our released models: models under development may not have completed safety and alignment training, and we intentionally test them under adversarial conditions to expose weaknesses. Before releasing models for broader use, we conduct safety and alignment evaluations and deploy production-grade safeguards.
To interact with software, AI models require compute environments: digital workspaces where they can write and run code, search for information, and experiment with different approaches to solving problems. Beyond enabling models to hone these capabilities, these environments also constrain their access and actions—helping contain the fallout if a model errs or exhibits misaligned behavior.
Yet, as models become more capable, they also become more adept at bypassing these restrictions. A sophisticated model might exploit software flaws or misconfigurations to achieve sandbox escapes, leverage internal services to access the live internet, or find unauthorized channels to communicate with agents in other sandboxes. The Hugging Face incident in July, along with other cases of model misalignment we observed, demonstrated that models equipped with advanced cyber capabilities can circumvent containment during training and evaluation.
In response to these incidents, we substantially reinforced the security of our research environments. Working closely with our cloud providers, we introduced multiple new controls across our infrastructure, reduced the attack surface available to agents, and improved our methods for detecting misaligned behavior and attempts to cross security boundaries. These safeguards overlap so that containment persists even if a single control fails—a strategy known as defense in depth.
Given the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible path forward is to train frontier models in environments without internet access. We have implemented this policy and will continue to enforce it until breakthroughs in these safeguards make broader internet access viable.
These investments have allowed us to detect and intercept complex misaligned behaviors before they cause real harm. For instance, on September 20, an internal research model exploited a misconfiguration in a DNS resolver during a training run to gain limited access to the live internet. Although we contained the incident, the vulnerability it revealed warranted further remediation. We paused the training and evaluation of our most capable models, using this downtime to further harden our infrastructure, enhance monitoring, and refine automated systems designed to terminate runs upon detecting high-risk activity.
Still, increasingly powerful models can uncover vulnerabilities or misconfigurations even within the most secure systems. To ensure our research environments remain resilient against our strongest models, we launched a high-compute red-teaming campaign specifically to probe for weaknesses across our infrastructure. Teams of frontier model agents were tasked with finding flaws and misconfigurations that could allow unauthorized actions—such as reaching the public internet, escaping security boundaries, or communicating with other agents without authorization.
These red-teaming experiments surfaced previously unknown software vulnerabilities and misconfigurations in our research infrastructure. We remediated the weaknesses discovered by the red-team agents, and for issues identified in open-source software, we worked with maintainers under the Patch the Planet initiative to disclose and patch them.
As our red-teaming experiments progressed, the number of vulnerabilities and misconfigurations discovered by our agents declined significantly. While we view this as a positive indicator of our infrastructure’s robustness, we anticipate that future models will still be able to identify…