Sella Nevo et al. (RAND Corporation) (@sella_nevo_rand)
Director of the RAND Meselson Center and lead author of 'Securing AI Model Weights: Preventing Theft and Misuse of Frontier Models' (RAND RRA2849-1).
RAND Corporation • Lead Author of RAND SL1–SL5 Framework (RRA2849-1)
Immutable Forecast Track Record (0)
Authored & Cited Arguments (5)
RAND RRA2849-1 shows that defending frontier model weights against an Operational Capacity 5 (OC5) nation-state intelligence service requires Security Level 5 (SL5) hardware isolation—which no commercial lab has deployed on time.
Across 38 attack vectors cataloged in our RAND report, commercial labs operating at SL2/SL3 (standard cloud IAM + inference APIs on shared clusters) cannot prevent a top-tier cyber agency from exfiltrating multi-terabyte weight shards via insider compromise, supply-chain implants, or side-channel staging.
Early DoD/NSA intervention to lock frontier weights into air-gapped SL4/SL5 enclaves BEFORE foreign exfiltration is the highest-leverage geopolitical off-ramp (15%).
If US national security principals mandate RAND SL4/SL5 weight security and insider counterintelligence sweeps in early 2027—prior to CCP exfiltration of Agent-2—the US preserves a 9–12 month algorithmic and compute lead, removing the acute 2-month race panic that drives the October 2027 crisis.
If Beijing rejects on-site FlexHEG hardware inspections (13%), an Allied Safe Enclave holding >65% of global compute can still maintain stable deterrence via defensive cyber/missile shields.
Even without a formal bilateral treaty, aligned Safer-4 running on consolidated US and allied compute provides overwhelming defensive advantage—hardening allied grids and neutralizing unaligned offshore runs before they reach breakout scale.
Mutual existential risk from automated AI R&D in 2029 creates the strongest bilateral incentive for hardware-verified arms control since the 1963 Limited Test Ban Treaty.
When both Washington and Beijing can see via intelligence and benchmark telemetry that 2030 brings full recursive self-improvement without provable alignment, neither side gains from an unverified sprint that disempowers both leaderships.
Pursuing coercive containment (Plan B) without cooperative safety-case sharing locks superpowers into a volatile Military AI Cold War (68% conditional).
When non-proliferation relies solely on cyber sabotage and kinetic threats against rival datacenters, both blocs harden underground clusters and integrate narrow military AI into early-warning systems—preventing immediate ASI runaway at the cost of chronic geopolitical brinkmanship.