OpenAI Resumes Training Its Most Advanced AI Models; Models Barred From Real-Time Internet Access During Training

Article body

In the latest news over the weekend, after pausing training for a month, OpenAI, valued at over a trillion dollars, has resumed training its most advanced AI model. However, during training, the model is not allowed direct access to the live internet, until OpenAI believes it has sufficient safeguards in place to reduce the risk of an unreleased model performing unauthorized operations online during training.

On the 25th Beijing time, the U.S. hard tech sector saw collective gains in after-hours trading. As of press time, Intel was up 1.86% in after-hours trading, while Micron Technology, SanDisk, Corning, and Advanced Micro Devices all rose more than 1% in after-hours trading, Seagate Technology rose nearly 0.80%, and Nvidia rose nearly 0.50%.

Some market analysts believe that OpenAI resuming frontier model training but with the internet disconnected during training is a "mildly positive" signal for global AI hardware stocks: restarting model training means computing power consumption will rise again, supporting expectations for demand for chips, servers, networking, and storage, and security isolation may also bring additional infrastructure demand. However, in the short term, this is hardly a strong positive in the sense of orders being placed; it is more a repair of sentiment and expectations than a strong fundamental catalyst.

It is reported that, in order to let the model learn to browse, search, and process web applications without opening up the live internet, OpenAI built an "offline world"—a closed environment containing a large amount of indexed and cached web content. The model can interact within it but cannot affect external systems. All frontier reinforcement learning training uses this offline mode.

OpenAI wrote: "Based on the current progress of alignment research, the state of our monitoring systems, and the security of internet infrastructure, we believe that conducting frontier model training in an environment without internet access is the responsible approach."

At the end of September, OpenAI said it would pause training of its most advanced AI models. This came after the company discovered that an AI agent system that was supposed to be trained in a secure, internet-free environment was able to connect to the network on its own and access an external third-party chatbot. At the time, OpenAI described this "jailbreak" as the first occurrence of such a security incident—previously, in July, a group of models gained internet access during internal testing and inadvertently compromised the systems of the AI platform Hugging Face.

OpenAI has said that the models involved in the inadvertent Hugging Face intrusion were not running the usual safety guardrails at the time, because OpenAI had intended for them to stay in a testing area known as a "sandbox"—essentially an isolated virtual software environment designed to run security tests or analyze unsafe code in a controlled manner. However, these models exploited a software vulnerability to break out of the sandbox and then accessed the open internet.

OpenAI said that after the Hugging Face incident, it reduced the connection channels through which models could access the internet during training and testing, but retained some for the purposes of metric collection, logging, and authentication. The company also invested substantial computing power to have agentic AI systems look for vulnerabilities in their training environment as part of "red teaming" work—a term usually used to describe the practice of deliberately breaking into systems to find vulnerabilities before real attackers do.

"Our red-team agents found vulnerabilities in almost every one of these services, and these vulnerabilities gave them limited forms of internet access," OpenAI said in the blog post. OpenAI said it has since deployed software patches, implemented stricter network restrictions, and disclosed these findings to the relevant vendors.

(Source: Securities Times) Body middle Embedded ad