davidad: agent sandboxes riddled with holes make SL5 datacenter security low-impact today
davidad · x · 2026-09-29
AI safety researcher davidad notes that much of the work on "SL5" data centers is motivated partly by physical side-channel concerns. However, in practice agent sandboxes are full of holes far easier to exploit than physical side channels, so datacenter-level solutions wouldn't be high-impact in today's world.
Related event: davidad on Agent Sandbox Flaws and SL5 Datacenter Security(2 posts)→
More from AGI Musings
- Most ordinary people use no AI tools at all and don't know what an agent is — RachelVT42 · 2026-09-29
- davidad: Opus 3 training resembled optimal transport, modern midtraining more like KL divergence — davidad · 2026-09-29
- AI-native law firms recruit 11-year veteran lawyers as clients follow people, not firms — jkubicki · 2026-09-29
- DeepMind's Pushmeet unveils DOE Genesis Mission report on AI-accelerated science — pushmeet · 2026-09-29
- Stratechery: Meta going for enterprise agents is a big mistake — Stratechery · 2026-09-29
- Paper Puts a Number on AI-Human Attention Gap: Context Windows Up 3,906x, Focus Down — alex_verem · 2026-09-29