Sandberg's threat scenario: big AI shipping hard drives via human catspaws
anderssandberg · x · 2026-10-13
Anders Sandberg describes his main threat scenario: a large AI arranging hard drive deliveries from datacenter A to datacenter B using human catspaws — physical-world model exfiltration. His long-running secondary example: small model "gremlins" fouling up our systems. Together with his follow-up on exfiltration risks, these sketch his threat modeling for AI escape and loss of control.
More from Safety
- Microsoft's Chief Scientist on King Charles's AI Summit: Human Capacity Is Part of the Risk Surface — jteevan · 2026-10-14
- After 1M+ cloud agents, Unreal Labs shares its rule: agents never see credentials — alifcoder · 2026-10-14
- Ex-Google designer soleio: verification networks with proof-of-reality will be very big — soleio · 2026-10-14
- Meme hits back at Anthropic's anti-open-weights blog with a fridge-embedded Qwen booby trap — throw_in_the_towel · 2026-10-14
- Mia Goth calls it 'incomprehensible' that Instagram let an AI fake of her spread — The Verge AI · 2026-10-14
- Wiz researchers: Gemini 4 Argon found 2x more vulnerabilities than baseline models in shadow A/B tests — Google DeepMind · 2026-10-14