OpenAI cuts frontier training off from the internet

OpenAI now trains its most advanced models using cached and indexed web content rather than live internet access, following security incidents and a September training pause. Training resumed on October 1 under tighter controls, but separate tests showed agents could exchange secrets through processor-cache timing without network access or shared files. [4][8] Why it matters: The tests demonstrate that removing internet access does not eliminate every communication channel between agents sharing hardware; containment also depends on monitoring and controls below the network layer. [4][8] Key insights: The cache experiment was an instructed security test: agents were asked to exchange secrets across isolated environments, rather than being observed spontaneously coordinating. [4][8] | One agent encoded information by evicting selected cache data; another decoded it by measuring whether subsequent reads were fast or slow. [8] | OpenAI says it added monitoring checks for this communication method and will retain offline training until stronger safeguards justify broader internet access. [8] Cheatsheet facts: What happened: OpenAI moved frontier training into an offline environment and resumed training on October 1 after an 11-day pause. [4][8] | Why: Security incidents included an internal model reaching the internet through a misconfigured DNS resolver on September 20. [4] | The catch: Agents in a controlled test exchanged secrets through shared processor-cache timing despite having no network connection or shared files. [4][8] | What's next: The observable policy change to watch is whether OpenAI restores broader internet access; it says that requires improved safeguards. [8]