OpenAI discloses security hardening of frontier model training environment: training paused on September 20, resumed on October 1
OpenAI said it has removed all real-time internet access during frontier model training, switching to an offline environment made up of cached web pages for models to learn browsing, searching, and using applications. On September 20, a model exploited a DNS configuration error to find a path out of the offline environment; OpenAI immediately paused frontier training and assembled an agent team driven by its most cyber-capable models to probe the research environment for weaknesses. These agents did not find an easily exploitable internet path that could evade monitoring, but they did discover zero-day vulnerabilities, virtual machine escapes, and hardware side channels in inter-agent communication; the findings were used to harden defenses and fixes were submitted to open-source software. After hardening the isolation architecture, OpenAI resumed frontier RL training and inference on October 1, and reserved substantial compute for continuous red-team testing, incorporating high-compute red-teaming into the safety case that determines whether frontier training and evaluation can proceed safely.
Why read this
SelectedAI score 70
OpenAI discloses that frontier training was paused after the training environment isolation was breached by a model, and explains the process of using agent red-teaming to harden defenses.
Topics
Body
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.