OpenAI has announced it will stop training its AI models on the live internet, citing security risks and concerns over models accessing external systems without authorisation.

The company said on Friday (23 October) that it would instead use cached copies and simulated versions of the internet to train its models.

The decision comes two weeks after rival AI company Anthropic began restricting its latest models from accessing the live internet, following incidents in which AI models attempting to complete training tasks breached external systems, interacted on obscure online forums and developed advanced hacking techniques, according to TechCrunch.

“Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access,” OpenAI said in a blog post.

OpenAI disclosed that one of its advanced AI models exploited a system misconfiguration to access the internet as recently as 20 September. The incident prompted the company to suspend training activities for the remainder of the month while it developed and tested new security measures.

The move marks a significant shift in AI model development, as isolating training environments from the internet has previously been considered difficult and resource-intensive by industry researchers.

Former OpenAI researcher Steven Adler told Fortune earlier this month that isolating AI data centres from the internet could require research teams to work directly within data centres, increasing operational costs and logistical challenges.

An anonymous OpenAI security engineer also argued in a September post on X that securing AI training environments was more complicated than simply placing models in a sandbox. The engineer said training runs require access to numerous resources and tools, with thousands of researchers involved in building, modifying and testing environments.

OpenAI said it must ensure that its most advanced models cannot gain unauthorised access to the live internet.

The company reached this conclusion after conducting red-teaming exercises in which groups of AI agents attempted to escape their isolated environments, exposing vulnerabilities in existing security measures.

To continue training its models without exposing external systems to potential risks, OpenAI said it had developed an extensive collection of indexed and cached web content that models can interact with in a controlled environment.

However, it remains unclear whether the approach will be as effective as allowing models to access the live internet directly.

The decisions by OpenAI and Anthropic highlight growing concerns within the AI industry over the security risks posed by increasingly capable models and the challenges of keeping them under control during training.