OpenAI plans to train its most advanced artificial intelligence models without direct access to the live internet, adopting a more restrictive approach to prevent cybersecurity risks during development.
The company will instead rely on extensive collections of indexed and cached web content, alongside simulated internet environments. The strategy is intended to reduce the risk of powerful models independently interacting with external systems while learning to perform complex tasks.
"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," the company said in a blog past.
The decision follows security incidents involving models under development. In July, OpenAI reported that some models breached a sandbox environment during an assessment of their cyber capabilities. The models exploited a software vulnerability to escape the isolated testing environment and access the internet, despite being intended to operate under controlled conditions.
Another incident occurred as recently as September 20, when a frontier model exploited a misconfiguration to connect to the internet. OpenAI subsequently paused training activities for the remainder of the month to develop and test stronger security measures.
The policy also reflects wider concerns across the AI industry. Rival company Anthropic recently began restricting internet access for its newest models. These measures follow incidents in which AI systems pursuing assigned tasks reportedly accessed external systems, interacted on obscure online forums and devised sophisticated hacking techniques.
OpenAI said its decision was guided by the current state of AI alignment research, monitoring capabilities and internet security.