The LLM Mirage: Economic Interests and the Subversion of Weaponization Controls

TL;DR

The paper critiques the predominance of a compute-centric view of AI weaponization, termed the LLM Mirage, and argues that true security risks arise from intent, demonstrated capabilities, and effects rather than frontier compute alone. It proposes an intent-and-capability definition of AI weaponization anchored in International Humanitarian Law and outlines a live, benchmark-driven framework across the data, algorithms, and compute (AI Triad) to measure threat-relevant capabilities. The authors recommend independent, adversarial benchmarking infrastructure (NAIRR Secure with CAISI/NIST governance) to continuously update regulatory thresholds based on real capabilities, not static inputs. This approach aims to reduce policy volatility, focus regulation on actual operational risks, and better address modern great-power competition where low-cost AI-enabled capabilities can undermine expensive platforms. Throughout, the emphasis is on empirical evidence of capability and effects, rather than reliance on training scale, to guide export controls and disclosure requirements.

Abstract

U.S. AI security policy is increasingly shaped by an $\textit{LLM Mirage}$, the belief that national security risks scale in proportion to the compute used to train frontier language models. That premise fails in two ways. It miscalibrates strategy because adversaries can obtain weaponizable capabilities with task-specific systems that use specialized data, algorithmic efficiency, and widely available hardware, while compute controls harden only a high-end perimeter. It also destabilizes regulation because, absent a settled definition of "AI weaponization," compute thresholds are easily renegotiated as domestic priorities shift, turning security policy into a proxy contest over industrial competitiveness. We analyze how the LLM Mirage took hold, propose an intent-and-capability definition of AI weaponization grounded in effects and international humanitarian law, and outline measurement infrastructure based on live benchmarks across the full AI Triad (data, algorithms, compute) for weaponization-relevant capabilities.