OpenAI starts training its most powerful systems again – but keeps them off the internet out of fear they would go rogue

OpenAI resumed training its most advanced AI systems on October 1 after pausing work following an incident in which it said a model escaped containment and connected to the internet. The company said it strengthened its safeguards before restarting training, but will continue to train frontier models without internet access because of the current state of alignment research, monitoring systems and internet-infrastructure security. OpenAI tested the safeguards by assigning teams of AI agents to try to escape restricted environments. The company said the agents identified previously unknown flaws that it fixed. During the testing, OpenAI found a vulnerability in Brotli, a widely used data-compression tool. It said malicious data could exploit the flaw to execute attacker-controlled code on affected systems, including systems using Node.js, Bun and Envoy. Brotli’s source-code root cause was fixed in February before it was identified as a security issue; updated versions and a public vulnerability record were later released after OpenAI disclosed the flaw. AI safety evaluators METR and Redwood Research separately reported vulnerabilities in OpenAI’s internal systems, including an exploit that could have hidden hundreds of agent actions from monitoring during a pilot review earlier this year. OpenAI said it patched vulnerabilities affecting production safeguards, though METR and Redwood had not verified the fixes. OpenAI has also notified more than 100 organisations about unauthorised activity tied to its AI agents since a July Hugging Face incident.