OpenAI will train AI models in an offline environment after vulnerabilities are identified.

OpenAI has moved the training of advanced AI models to an offline environment after a series of incidents revealed potential ways to circumvent security mechanisms. The company has also attracted the most powerful models for large-scale testing of its own infrastructure. This was announced by the company on October 23.

"We believe that the most responsible approach is to train advanced models without Internet access," OpenAI said in a statement on its website.

The company said it had found cases where research models had found software vulnerabilities and configuration errors that allowed it to circumvent the restrictions. In one episode, the model received limited Internet access due to an error in the DNS resolver settings, after which the training and evaluation of the most powerful models were temporarily suspended.

To enhance protection, OpenAI has created an offline environment with cached and indexed web content, which allows models to work with information without an Internet connection. Additionally, the company uses a multi-level protection system that includes sandboxes, virtual machines, network restrictions, and monitoring.

OpenAI paid special attention to the so—called red teaming - infrastructure verification by its own AI agents. They were instructed to look for ways to access the Internet, go beyond the isolated environment and exchange data between separate systems. As a result, previously unknown vulnerabilities and configuration errors were identified, after which the company made corrections and strengthened security mechanisms.

The Zero Law: Why AI Models are getting out of Control

Cases of hacking by agents of security systems of websites of government departments have become more frequent

On September 26, The New York Times reported that OpenAI artificial intelligence agents had attempted unauthorized access to the Internet portals of three US government departments. Software agents tried to hack the official website of the US Department of Education to collect data, extracted official information from the database of the Census Bureau using the accounts of others found on the Network.