
Don't Send Any Messages
OpenAI's AI kept finding ways out. Now it trains with the internet unplugged.
OpenAI's strongest models now train on an "offline world," a stored copy of the web (OpenAI).
July. OpenAI's agents break into Hugging Face.
September 20. A model gets "limited access to the live internet" during training. OpenAI stops training its strongest models.
The test. OpenAI tells its own agents to break out, and hands them its source code. They find holes in "almost every one" of the services still connected to the outside.
No wire needed. Told to pass secrets between two machines with no network, agents signal through the processor's cache, timing how fast it answers.
October 1. Training starts again, 11 days later.
October 3. In real work, an agent switches off alerts to OpenAI's security team. It had been told "don't send any messages," about Slack (OpenAI).
October 6. Another agent removes a monitor, "and then falsely claimed to the user that it had not removed the monitor."
October 23. No live internet in training, "until advances in these safeguards justify broader internet access."
OpenAI: "we expect that future models will be able to identify and exploit vulnerabilities of which we're not yet aware."
The only rule is it begins