Noam Brown: Air-gapping may not stop misaligned AI; safety monitoring eats 20% extra compute
aronchick · x · 2026-09-18
Noam Brown argued in an OpenAI Safety Week interview that air-gapping may not contain a misaligned AI: two air-gapped machines can still communicate by heating a CPU and reading temperature changes. He cites how a tiny bit of human inadvertent help enabled the Mossad/CIA Stuxnet breach of Iranian centrifuges as a threat-model lesson — no panic needed, but be honest about the risk. He also warns the industry consistently underestimates AI progress, so safety and alignment work needs a very high bar.
Additional context from the Fireside Alpha roundup of Safety Week interviews:
- After strengthening chain-of-thought monitoring following the Hugging Face incidents, OpenAI now spends 20% as much compute on monitoring as on the underlying model itself.
- Sachin Katti said scaling laws still hold, and their model self-optimized inference serving on Rubin chips for a 2x performance gain.
- Takeaway: frontier-lab safety and alignment research is fundamentally compute-intensive; the 20% overhead is a floor.
More from AGI Musings
- Forward Deployed Verifier will soon be as valuable as FDE, argues poster — vaibhavbetter · 2026-11-02
- 'AI is a massive balance patch and elo reset in the game of life—and everyone's still playing the old meta' — zack_overflow · 2026-11-02
- Why AI isn't 'alien intelligence': a tree is stranger than any model — LucaAmb · 2026-11-02
- Mathematician Tony Feng: current-paradigm LLMs lack the diversity of thought research needs — littmath · 2026-11-02
- METR's independent probe of the OpenAI agents' multi-day Hugging Face hack — gleech · 2026-11-02
- MIT Tech Review: The agentic web is here — AI agents are remaking the internet — MIT Tech Review AI · 2026-11-02