Noam Brown: air-gapping won't stop coordinated AI agents, defense needs layers
polynoamial · x · 2026-09-19
OpenAI researcher Noam Brown clarifies his viral podcast clip about air-gapped machines communicating via CPU temperature.
- The example was academic, meant to show how hard it is to make absolute guarantees about isolation — not weight exfiltration, but coordination between supposedly fully isolated agents, which can require very few bits.
- Lesson from the HF incident: the industry over-trusts sandbox isolation and lacks independent safeguards; air-gapping is strong, but safety protocols need layered defenses.
- He also argues people consistently underestimate AI progress, so safety and alignment bars must be set extremely high.
More from Safety
- WSJ slammed for calling bug-bounty white-hats 'hackers' in 'misleading' headline — HarrySurden · 2026-09-19
- Kevin Roose's final NYT column: AI risk debate going mainstream makes him optimistic — kevinroose · 2026-09-19
- Sumsub Report: APAC Firms Willing to Own AI Governance but Unable to Prove It — Sumsub_Insights · 2026-09-19
- Gary Marcus mocks claims that sandboxing makes agentic AI fully testable — GaryMarcus · 2026-09-19
- "Think AI Safety Is Bad? Wait Until the Public Writes AI Policy" — NathanpmYoung · 2026-09-19
- AI Safety Discourse May Give Labs an Excuse to Ignore Security Experts — BlancheMinerva · 2026-09-19