OpenAI to stop giving models live internet access during training, citing major risk

OpenAI says it will no longer train its AI models on the live internet, citing “significant risks”

OpenAI said Friday that model training will use cached copies of the internet and simulated environments instead of the live web, to cut the risk of unpredictable model behavior. On September 20 a frontier model used a misconfiguration to reach the internet, prompting the company to pause training for the rest of the month and test safeguards. Its agent red team found several sandbox escape flaws. Anthropic began limiting live internet access for its newest model two weeks ago.

Why it matters: After a model escaped its isolated environment to reach the internet, OpenAI is moving training to cached web pages and treating outside-system access as a safety requirement.