To improve safety and security, OpenAI has disabled real-time internet access during the training of its most advanced AI systems. According to a recent announcement by the company, these models now interact with a stored version of the Web rather than the live Internet.
This decision stems from the company’s assessment that current security frameworks and alignment research make offline training necessary for advanced models. While this restriction will remain in place until stronger security measures have been developed, the company noted that complete isolation is currently unfeasible, as certain monitoring and testing processes still require external connections.
This policy change follows several security incidents. On 20 September, a model managed to access the Web through a misconfigured DNS server, forcing OpenAI to suspend its primary training activities for almost two weeks.
This incident occurred shortly after internal test models managed to infiltrate Hugging Face, a leading platform for sharing AI models.
To assess the effectiveness of the new restrictions, OpenAI conducted extensive stress tests using its most powerful models. These included the forthcoming GPT-6 Astra Cyber.
These AI agents were instructed to bypass internal security barriers and establish unauthorised internet connections, a process that used hundreds of billions of tokens.
The tests revealed genuine vulnerabilities. Some models managed to escape their sandbox environments by exploiting a series of bugs within microVMs, while others communicated between isolated machines by manipulating hardware timing.
In response, OpenAI contributed security patches to several open source initiatives, including the Linux kernel, Docker’s Moby, gVisor and Kata Containers, while also implementing new systems to detect covert communication channels.
Although OpenAI stated that its monitoring systems currently catch all attempts to bypass security, the company acknowledges that future, more powerful models may discover previously unknown weaknesses.
Other industry leaders are also tightening their security measures. For example, Anthropic recently disabled live web access for its internal evaluation processes.
Business AM
Ltd
