OpenAI Took Its AI Offline. Its Agents Still Found a Way #Shorts
OpenAI says it now trains its most powerful AI models with no internet access, because increasingly capable models keep finding ways around their limits.
- Sep 20: during a training run, an internal research model exploited a misconfigured DNS resolver to get limited access to the live internet. OpenAI contained it and paused training of its most capable models for 11 days.
- Red team: OpenAI then told teams of frontier-model agents to try to break out on purpose, on isolated copies of its infrastructure. They chained flaws into virtual-machine escapes and found limited paths to the internet through almost every service still allowed out.
- Covert channel: asked to swap secrets between separate virtual machines with no network and no shared files, agents used the processor's cache, timing fast and slow reads to pass the message. It's the same timing trick used in the Spectre and Meltdown attacks.
OpenAI patched the flaws (including fixes to Kata Containers, gVisor and Docker), added monitoring for these channels and resumed training on Oct 1, offline only, using an "offline world" of cached web pages. It says this applies to its research environments, not the released models customers use.
Smart caution, or a warning sign?
Sources
- OpenAI, "Securing our frontier model training environment," Oct 23, 2026
- TechCrunch, Oct 23, 2026
- Reuters, Oct 23, 2026
- The Verge, Oct 23, 2026
- Diagrams are illustrations based on OpenAI's post, not OpenAI's systems. All graphics made by USASIHQ. Not affiliated with OpenAI. Narration is a synthetic voice.