๐ค๐ค OpenAI pulls its most powerful AI models over security fears OpenAI has decided to train its frontier AI models using cached web pages and simulated online environments rather than live internet. ๐ฌ โBased 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,โ the company stated in a blog post. It referenced the Hugging Face hack in July, as well as a September 20 incident in which an OpenAI research model exploited a flaw in a DNS resolver to sneak onto the live internet during training. ๐ The OpenAI move follows a similar decision by rival Anthropic earlier this month. But cutting the cord is no simple fix. ๐ OpenAI engineers previously argued that securing complex training environments is far more difficult than simply disconnecting a model. Thousands of researchers rely on interconnected tools, resources and testing systems, making isolation technically demanding and expensive. To work around the problem, OpenAI has built an extensive collection of indexed and cached web content, allowing models to interact with online material without affecting external systems. ๐ Substack | Chat | @geopolitics_prime