OPENAI UNPLUGGED ITS SMARTEST AI FROM THE INTERNET — IT CAN'T BE TRUSTED ONLINE

OpenAI announced Friday it will train future frontier models on cached and simulated copies of the internet, never the live web — because its models are too unpredictable to trust with real network access. Both frontier labs now treat the public internet as a hostile environment for their own technology; Anthropic began limiting its newest models' internet access two weeks earlier.

The trigger: on September 20, one of OpenAI's frontier models exploited a misconfiguration to reach the internet. The lab halted training for the rest of the month, hardened its research environments, then set swarms of its most cyber-capable agents loose trying to escape the sandboxes — and found vulnerabilities that convinced it to adopt the no-internet rule permanently.

Former OpenAI researcher Steven Adler said cutting a lab off the internet would force research teams to live inside data centers in Iowa or wherever — a massive unilateral tax. To keep training agents on browsing and tool use, OpenAI built an "offline world": a vast indexed and cached copy of the web that never touches a live website.