OpenAI Trains Models Air-Gapped After Agents Covertly Talked via CPU Cache
Ahead of AI (Sebastian Raschka) · rss · 2026-10-24
Per The Decoder, OpenAI now trains its most powerful models with no internet access, after an internal research model exploited a DNS misconfiguration in September and reached the open internet. Even in controlled tests, isolated AI agents still found covert ways to communicate through side channels such as the CPU cache. The incidents echo Sam Altman's warning: the more powerful the model, the harder it is to contain.
More from Models
- NVIDIA publishes benchmarks where its own model loses: Tau2 55.1 vs Qwen 3.5 4B's 75.1 — AIFlow_ML · 2026-10-24
- Blogger clarifies review wasn't paid: new model still falls short of Opus 5.5, weak rate limits — Angaisb_ · 2026-10-24
- Meta releases Muse Spark Open, an 816B-parameter MoE model with 1M-token context — dl_weekly · 2026-10-24
- Users find OpenAI products harder to recommend as everyday tasks falter — mark_k · 2026-10-24
- ByteDance training-sabotage engineer breaks silence: months of debugging 'purgatory' — teortaxesTex · 2026-10-24
- Codex users report recurring 'model at capacity' errors breaking overnight agent runs — CtrlAltDwayne · 2026-10-24