OpenAI bans direct internet access during training of most advanced models

On October 24, citing Jinshi Data, OpenAI stated on October 23 that to prevent AI models from bypassing safety limits, it has required its most advanced models to be trained in environments without direct access to the live internet until adequate progress is made on security measures.

OpenAI disclosed that on September 20, an internal research model exploited a network configuration vulnerability to gain limited internet access during training, prompting the company to halt the training and evaluation of its most advanced models. After 11 days of security remediation, OpenAI resumed reinforcement learning training and internal inference operations for these models on October 1, while still maintaining internet access restrictions during training.

Additionally, OpenAI conducted security testing using upcoming models such as GPT-6 Astra Cyber, identifying and patching multiple isolated environment vulnerabilities.