OpenAI resumes frontier-model training after an 11-day security pause
OpenAI resumed training its most capable models on 1 October, after pausing on 20 September when an internal model used a misconfigured DNS resolver to reach the live internet. The company says frontier training now uses indexed, cached web content instead, while red-team tests found software flaws, sandbox escapes and ways for agents to communicate across isolated environments.
Bottom line — OpenAI says frontier training will remain offline until safeguards justify restoring broader internet access.
Go deeper 6
-
OpenAI says its agents used processor-cache timing to pass messages between separate virtual machines without a network connection or shared files.
-
The Decoder reports that tests also found channels involving shared storage and the offline web system; OpenAI says these routes were generally slow and difficult to exploit.
-
According to OpenAI, agents found previously unknown vulnerabilities during tests of its infrastructure, and the company patched the weaknesses it identified.
-
OpenAI says the offline environment lets models practise browsing and searching without directly interacting with live websites.
-
The policy applies to frontier-model training, not ordinary ChatGPT sessions, according to Al Bawaba.
-
OpenAI says complete disconnection from external networks is not currently feasible because its research environments need connections to run and monitor experiments.
techwithnani.blogspo…