OpenAI had teams of its own frontier agents attack copies of its training infrastructure after a model reached the internet during a September 20 training run, and the agents found ways to break out of the virtual machines meant to contain them. Once in control of a host, they found bugs giving limited internet access in almost every service OpenAI still lets connect out, including metrics, logging, and login systems. Agents in separate virtual machines also passed messages to each other by timing how fast the processor’s cache returned data, the same trick behind the Spectre and Meltdown attacks. OpenAI says it patched the problems and restarted frontier training on October 1, after an 11-day pause, with those models cut off from the internet entirely.
