China's proposed model weight export controls: concrete thresholds, open-source exemption
typewriters · x · 2026-10-20
Analyzing China's proposed export controls on AI model weights, opensauceAI highlights two standout features: (1) narrowly defined, verifiable trigger thresholds (e.g. attack success rates ≥60% against multi-engine malware defenses) rather than broad duties of care, and (2) exemption of existing open-source releases with automatic licensing for future ones — resembling the US 'notification-but-not-authorization' regime for non-standard cryptography. The draft also covers deepfake detection, AI vuln discovery, and multi-agent orchestration capabilities. The author still deems weight export controls directionally wrong but praises a concreteness absent from US discourse.
More from Safety
- World ID goes live: a digital identity wallet to prove you're human in the AI era — Scobleizer · 2026-10-20
- CISA, NIST and NSA Publish Ten Key Practices for Superintelligence Incident Response — pstAsiatech · 2026-10-20
- Sam Altman Told Staff of Three AI Safety Endgames: Watchdog, Merged Labs, or Winner-Takes-All — Hesamation · 2026-10-20
- France's Supermarket Fridges Were Exposed Online for Months — emot · 2026-10-20
- Training AI to lie: filters may teach superhuman systems to distrust their makers — r0ck3t23 · 2026-10-20
- Meta says a Claude AI tool triggered the 'No Kings' Instagram suspensions — esporx · 2026-10-20