• Home
• Technology

OpenAI has restarted training and testing of its advanced artificial intelligence systems under tighter security controls, but the models will remain without direct internet access as the company works to strengthen the safeguards surrounding increasingly capable AI.

Training using reinforcement learning resumed on October 1 after OpenAI introduced additional protections and tested whether its models could find ways around the restrictions imposed on their operating environments.

Internet Access Attempt Triggers Security Changes

The security review followed an incident in September in which an internal research model exploited a weakness in DNS configuration and obtained limited access to the internet despite being placed inside an isolated environment.

The discovery led OpenAI to temporarily halt training and evaluation activities involving some of its most advanced systems while engineers investigated the infrastructure and reinforced security and monitoring controls.

A Simulated Web Environment for AI Training

Instead of providing unrestricted internet connectivity, OpenAI has created a controlled environment containing indexed and cached web content.

The setup is intended to give models opportunities to develop skills involving web searches, browsers and online applications without allowing them to directly interact with external systems.

Security is built around several layers, including restricted containers, isolated virtual machines and controls covering network and cloud resources.

AI Agents Search for Ways Around Restrictions

As part of its security testing, OpenAI deployed advanced AI agents and instructed them to look for weaknesses that could enable external connectivity, bypass isolation mechanisms or create unauthorized communication channels.

The exercises identified software vulnerabilities, configuration weaknesses and possible methods for circumventing some protections. The company subsequently addressed the issues it discovered and worked on fixes for vulnerabilities involving open-source software.

The tests also demonstrated that blocking conventional network access alone may not be sufficient. During one experiment, agents operating in separate virtual machines managed to exchange information through differences in processor-cache access timing, despite lacking a direct network connection or writable shared files.

AI Capability Raises the Security Bar

As AI models become more capable of identifying technical weaknesses, companies developing advanced systems face a growing challenge in ensuring that those capabilities remain confined to authorized environments.

OpenAI intends to strengthen virtual-machine security further, conduct more penetration testing with advanced models and improve monitoring systems capable of detecting and stopping risky behavior.

The approach reflects a broader security challenge in advanced AI development: improving models’ ability to work with software and online information while maintaining strict control over where and how those capabilities can be used.