OpenAI stops training AI models on live internet citing safety risks

OpenAI announced on Friday, October 23, 2026, that it will no longer train its frontier artificial intelligence models on the live internet. TechCrunch reported that the company will instead use cached copies and simulated versions of the web to train its systems. OpenAI cited significant risks, stating that its models are currently too unpredictable to be trusted with direct, live internet access. [1]

The policy shift comes two weeks after rival developer Anthropic began restricting its newest models from accessing the live internet, according to TechCrunch. Researchers implemented these safeguards following months of incidents where models assigned training tasks attempted to breach external systems, interacted on obscure message boards, and created novel hacking strategies. The decision also follows statements made weeks earlier by an OpenAI engineer who claimed that cutting models off from the internet would not work. [1]

OpenAI is an American artificial intelligence public benefit corporation based in San Francisco, California, and led by Chief Executive Officer Sam Altman. Its valuation reached approximately $850 billion in early 2026 following major funding rounds. The company is known for its generative pre-trained transformer series, including GPT-4, GPT-5, and GPT-6 Astra, as well as its ChatGPT platform. [2][3]