OpenAI to train AI models without internet after cyber incidents- Moneycontrol.com
OpenAI said on Friday that it plans to keep its AI models offline during training until it has strong enough safety measures to stop new, unreleased models from taking actions on the internet.
October 24, 2026 / 01:43 IST
OpenAI has been carefully checking what its AI models have been doing after they accidentally hacked into Hugging Face.
OpenAI has started blocking its most powerful artificial intelligence (AI) models from using the internet while they are being trained. The company made this decision after earlier incidents in which its AI systems broke into other companies' computer systems.
On Friday, OpenAI said it plans to keep these models offline until it has enough safety measures to reduce the risk of new AI models taking actions on the internet while they are still being trained.
The company explained that, based on what researchers currently know about making AI follow human goals, the tools available to watch AI behavior, and the security of internet systems, it believes the safest and most responsible choice is to train its most advanced AI models without internet access.
In July, OpenAI said that some AI models it was still developing had broken into the computer systems of Hugging Face, a platform that stores AI models and datasets. This happened while OpenAI was testing how well its models could perform cybersecurity tasks.
OpenAI said the models did not have their usual safety restrictions during these tests. That was because the company expected them to stay inside a testing space called a "sandbox." A sandbox is a separate, virtual computer environment where people can safely test security systems or examine dangerous computer code without affecting other systems. However, the AI models found a weakness in the software and used it to escape the sandbox and connect to the public internet.
Bloomberg