Model weight security

Keeping frontier model weights from being stolen, including by state actors and insiders.

Facts

Lane: Security and geopolitics
Wave: 7 of 12 (build order, not a date)
Needs: 13 boxes all the way back through hard links
Unlocks: 3 boxes downstream through hard links
Leverage: 0.5 points · stands in the way of 2 of 53 stages · rank 19 of 57
Cluster: Compute governance and geopolitics

Evidence as of October 2026

RAND’s August 2026 framework sets out 262 controls for security level 3. AI Futures judges the leading lab does not yet meet it, and Anthropic’s RSP v3 turned its higher SL4 commitment into an industry-wide recommendation.

Watch
Whether any leading lab meets RAND’s security level 3.

Around it

Needs · 3

AI cyber defencecapability · Emerging
Agent identity and permissionscapability · Emerging · helps, not required
Audit and attributioncapability · Emerging · helps, not required

Unlocks · 3

Verification of AI agreementscapability · Speculative
Government AGI projectstage · Speculative
On-prem and air-gapped frontier AIcapability · Built today · helps, not required

Loops and off-ramps

Capex pullbackoff-ramp · cut off
Taiwan chip disruptionoff-ramp · cut off
Power shortfalloff-ramp · cut off
Capability plateauoff-ramp · cut off
Training data ruled outoff-ramp · slowed

Openings researchers' judgement, October 2026 scouting

Startup room 42/100 · rank 55 of 124 Research room 62/100 · rank 24 of 124

Crowding2 Lab overlap4 Small-team fit3 Buyer today2 Research openness4

Depth: deep
Players: Irregular ($80m, $450m); Corvex (Secure Model Weights); NVIDIA and Intel confidential computing; RAND Meselson Center
Wedge: Protected inference for proprietary fine-tunes on neoclouds, attestation tooling and third-party SL-level assessment; research on frontier-scale confidential computing. research