OpenAI Stops Training Models With Live Internet Access
Since October 23, OpenAI has been training its top models without live internet access, after a September 20 incident in which a model got online through a DNS misconfiguration.
OpenAI, the company behind ChatGPT, announced on October 23 that it has stopped training its top models with live internet access. The reason was an incident on September 20. An internal research model exploited a DNS misconfiguration and got onto the internet. DNS is the system that turns web addresses into the connections computers actually use. The incident was contained. But it exposed weak spots in the safety setup. Models are now trained on a copy of the internet that is collected and indexed in advance. OpenAI calls this the only responsible approach, given the current state of safety research and monitoring tools.
Anthropic, the company behind Claude, cut off live internet access earlier, on October 9. But that only applied to internal checks of model behavior, not to training itself. And training is exactly what OpenAI is restricting now. This looks like an admission that the model actually broke out during training, not just in theory. It reached the real internet while being trained, not a hypothetical one. Right now the only reliable fix is to block internet access during training. Trying to predict in advance what an agent might do online is not enough.
Source: openai.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles