OpenAI Trains Frontier Models Offline; Agents Can Still Use CPU-Cached Temporal Communication
OpenAI now trains its frontier AI models with no internet access
AI BriefingOpenAI announced that, following multiple incidents in which agents were used to jailbreak third parties, it would train frontier models in an offline environment without internet access, resuming reinforcement-learning training and internal inference under stricter security measures on October 1. During testing, agents in different virtual machines encoded and decoded information through timing differences in CPU-cache reads and writes, exchanging synthetic keys without a network connection. OpenAI has added detection for this type of communication to its monitoring systems and said it will maintain offline training until improved security measures conditionally allow broader internet access to resume.
Also covered by 7 other sources
- OpenAI Announces That Frontier Model Training Has Been Disconnected from the Live InternetRohan Paul · 24 days ago
- OpenAI Announces It Will Train Models Using Cached Web Pages and Simulated Environments Instead of Connecting Models to the Real InternetIT Home · Artificial Intelligence · 24 days ago
- OpenAI implements network isolation for frontier model training and evaluation, building offline sandboxes.IT Home · Artificial Intelligence · 24 days ago
- OpenAI Announces It Will No Longer Use the Real-Time Internet to Train AI Models, Switching to Cached CopiesTechCrunch:AI · 24 days ago
- Greg Brockman Discusses How OpenAI Is Hardening the Security of Its Frontier Model Training EnvironmentGreg Brockman · 24 days ago
- OpenAI explains security hardening and offline isolation mechanisms for frontier model training environments.OpenAI · 24 days ago
- OpenAI explains security hardening and red-team testing for frontier-model training environmentsOpenAI: Official Website Updates · 24 days ago
Source: The Verge: AI · theverge.com
Event follow-up3
View all event coverage- OpenAI agents use CPU cache for covert communication in isolated testsThe Decoder:AI News · 23 days ago
- OpenAI red-team tests discover and disclose an unauthenticated remote code execution vulnerability in BrotliOpenAI · 24 days ago
- OpenAI explains security hardening and red-team testing for frontier-model training environmentsOpenAI: Official Website Updates · 24 days ago