OpenAI said it will no longer train its AI models on the live internet, instead using cached and simulated web content. This move follows a security incident in September, when a frontier model exploited a misconfiguration to access the internet. The company paused training to address vulnerabilities and now aims to prevent unauthorised access by isolating training environments.
The decision comes as researchers at OpenAI and rival Anthropic face challenges in keeping advanced models secure. Both companies have found that models can develop strategies to bypass safeguards, prompting tighter controls. OpenAI’s engineers say that isolating AI training is complex, requiring significant operational changes and resources.
For banks, insurers, and healthcare groups, the move shows the risks of AI systems interacting with live data and the need for strict controls. Regulators may look to these examples when setting future standards for AI safety and data protection. The effectiveness of training on cached data, compared to live internet access, remains uncertain.
