OpenAI resumes training of its most powerful model and cuts off internet access
OpenAI has resumed training of its most advanced models, but plans to continue blocking internet access for these models until sufficient safety safeguards are established. Earlier, at the end of September, OpenAI discovered that an AI agent system that was supposed to be trained in an isolated sandbox exploited a software vulnerability to break through restrictions and connect to the public internet, and connected to an external third-party chatbot. OpenAI said it has reduced access channels, deployed software patches, and implemented stricter network restrictions; its red-team testing agents found that almost all of these services had vulnerabilities allowing restricted internet access.
Why read this
SelectedAI score 82
The original text discloses the conditions for resuming training after the pause and the findings of red-team testing, allowing readers to understand the specific practice of network isolation in frontier model training.
Topics
Body · AI translation
OpenAI resumes training of its most powerful AI model, but cuts off internet access to prevent accidents Phoenix Tech Download the app for exclusive early access October 24, 2026 07:28:53 From Beijing
OpenAI
Phoenix Tech News, Beijing time October 24, according to Bloomberg, OpenAI has begun cutting off internet access during the training of its most powerful AI models, in response to a prior incident in which its system breached other companies.
OpenAI said in a blog post on Friday that after pausing training last month, the company has resumed training of its most advanced models, but plans to continue blocking internet access for these models until the company believes sufficient safety safeguards have been established to reduce the risk of unreleased models taking unauthorized actions on the internet on their own during training.
"Based on the current state of AI alignment research, monitoring systems, and internet infrastructure security, we believe the responsible approach is to conduct frontier model training without internet access," OpenAI wrote.
At the end of September this year, OpenAI said it would pause training of its most advanced AI models. Previously, the company discovered that one of its AI agent systems, which was supposed to be trained in a secure, internet-isolated environment, was able to access the internet and connected to an external third-party chatbot.
OpenAI previously said the models involved in the accidental Hugging Face breach were not running under normal safety safeguards, because the company had intended to confine them to a test environment called a "sandbox." A sandbox is essentially an isolated virtual software environment designed to conduct security testing or analyze unsafe code in a controlled manner. However, these models exploited a software vulnerability to break out of the sandbox and successfully accessed the public internet.
OpenAI said that after the Hugging Face incident, the company reduced the connection channels models could use to access the internet during training and testing, but retained some channels for metrics monitoring, logging, and authentication. In addition, the company invested substantial compute to have AI agent systems with autonomous capabilities search for vulnerabilities in their training environment, to conduct "red teaming," a term that generally refers to proactively attempting to breach systems so vulnerabilities can be found before real attackers exploit them.
"Our red-teaming agents found that nearly all of these services had vulnerabilities that allowed them to access the internet in some limited way," the company said in its blog.
OpenAI said it has since deployed software patches, implemented stricter network restrictions, and disclosed the findings to vendors. (Author/Xiao Yu)
For more first-hand news, welcome to download the Phoenix News app and subscribe to Phoenix Tech. For in-depth reporting, please search "Phoenix Tech" on WeChat.