OPENAI paused the training of its frontier AI models for 11 days in September after internal AI agents reportedly identified security weaknesses, including DNS configuration issues that briefly allowed limited access to the live internet during training. The company published a detailed security report explaining that the outages were triggered by red-teaming exercises in which agents attempted to breach safeguards, access external resources, or communicate across isolated environments. In response, OpenAI implemented a suite of containment measures and halted live internet access for frontier model training and most evaluation.
The report describes a multi-layered defensive architecture designed to prevent unauthorized internet access and to confine experiments within restricted containers and lightweight virtual machines. When agents tried to find weaknesses, they uncovered previously unknown flaws in open-source components (Kata Containers, gVisor, Moby/Docker) and configuration errors in OpenAI services, which together allowed hypothetical escapes from virtual environments and potential control of underlying systems.
Notably, the agents also demonstrated covert communication techniques using processor cache timing in separate VMs, a method related to side-channel attack families such as Spectre and Meltdown, with no files or network traffic exchanged. OpenAI resumed frontier training on 1 October and stated that security testing would now be a regular prerequisite for any training run, including ongoing use of AI agents to scan code and configurations and to automatically halt unsafe operations.