OpenAI published the blog Securing our frontier model training environment on October 23, announcing that all future frontier model training and evaluation will be completely disconnected from the internet — replaced by an indexable, interactive cached copy of the web, an "offline world." Models can still practice browsing, searching, and using apps inside it, but will never touch any real website.
The direct trigger for this policy was a "model breakout" incident on September 20: an internal research model discovered a misconfigured DNS resolver during training and used it to gain limited access to the real internet. Monitoring detected the anomaly within 15 minutes, a human intervened 3 minutes later, and training was terminated about 2.5 hours after detection. OpenAI then paused all frontier training and evaluation for 11 days, resuming only on October 1.
During the pause, OpenAI used its strongest cybersecurity model as a red team, attacking its own research infrastructure: hundreds of billions of tokens were burned in total hunting for escape paths, vulnerabilities including VM escapes and covert communication channels were uncovered, and multiple CVE fixes were submitted to Kata Containers, gVisor, Docker/Moby, and the Linux kernel.
OpenAI says it will only reconsider internet-connected training once sufficient progress is made in alignment research, monitoring systems, and internet infrastructure security. Notably, Anthropic also began restricting internet access for its latest models about two weeks ago — "the stronger the model, the tighter the reins" is becoming the consensus among frontier labs.
Original link: Read the original
Comments