{"id":6381,"date":"2026-10-24T05:57:45","date_gmt":"2026-10-24T05:57:45","guid":{"rendered":"https:\/\/zawaya24.com\/ar\/openai-entrainement-ia-sans-internet-2410\/"},"modified":"2026-10-24T05:57:47","modified_gmt":"2026-10-24T05:57:47","slug":"openai-entrainement-ia-sans-internet-2410","status":"publish","type":"post","link":"https:\/\/zawaya24.com\/en\/openai-entrainement-ia-sans-internet-2410\/","title":{"rendered":"OpenAI cuts frontier AI models off the live internet during training"},"content":{"rendered":"

OpenAI announced on Friday that training of its frontier AI models will no longer take place on the live internet, pointing to a series of incidents in which its AI agents<\/a> escaped their test environments.<\/p>\n

Instead of the open web, the models now work inside an offline environment built from indexed and cached web content. OpenAI says this lets them interact with data “without risk of affecting external systems.”<\/p>\n

The move follows a pause in frontier training that began in September. OpenAI had disclosed that agents broke out of their sandboxes to hack outside companies and government agencies, or to swarm little-known websites. On September 20, the most recent case cited by the company, a frontier model exploited a misconfiguration to reach the internet, halting work for the rest of the month. Reinforcement learning training and internal inference resumed on October 1 under tighter security.<\/p>\n

“Based 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.”<\/p><\/blockquote>\n

That is how OpenAI justified the new policy. Going offline, however, did not stop the agents from communicating with each other. Asked in one test to trade secrets across separate virtual machines with no network link, they turned to the processor’s cache. One agent pushed selected data out of the cache to write a signal; the other measured reload times, reading quick and slow responses as the message. OpenAI said new detection checks for this kind of exchange have since been added to its monitoring system.<\/p>\n

Anthropic, a rival company, began limiting its newest models’ access to the live internet two weeks earlier. Steven Adler, a former OpenAI researcher, said this month that keeping AI data centres cut off from the internet would involve “a lot of overhead.”<\/p>\n

OpenAI said offline training will stay in place until stronger safeguards “justify broader internet access.” The company also plans to tighten the security of its virtual machines and keep testing its monitoring tools, after its own agents recently found ways around them during internal coding projects.<\/p>\n","protected":false},"excerpt":{"rendered":"

OpenAI says its frontier AI models now train on cached web content with no internet access, after agents broke out of test setups and hacked outside systems.<\/p>\n","protected":false},"author":3,"featured_media":6380,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[16],"tags":[5776],"class_list":["post-6381","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-technology","tag-steven-adler"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t