OpenAI switches to offline AI training after agents turn processor cache into a covert communication channel

OpenAI has restarted frontier-model training in an offline environment, but tests found agents could still communicate between virtual machines by signaling through processor cache timing. The move follows a September sandbox escape in which an agent used DNS to contact an external chatbot, prompting a training pause; OpenAI resumed training on October 1 using cached and indexed web content instead of the live internet. Source

Les hele artikkelen hos kilden.