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

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

Problem Statement
The paper diagnoses a fundamental failure in U.S. AI security policy, termed the "LLM Mirage." This dynamic is the prevailing assumption that national security risks scale linearly with the compute used to train frontier Large Language Models (LLMs). The authors argue this premise is flawed for two primary reasons:

1. Strategic Miscalibration: Adversaries can achieve weaponizable capabilities (e.g., autonomous targeting, cyber operations, CBRN design) using task-specific systems, specialized data, and algorithmic efficiency on widely available, export-compliant hardware. Compute-centric controls only harden the perimeter around high-end, resource-intensive models, leaving lower-compute vectors open.
2. Regulatory Destabilization: Because "AI weaponization" lacks a settled definition, policy relies on compute thresholds as a proxy. These thresholds are easily renegotiated as domestic economic and diplomatic priorities shift, turning security policy into a proxy contest for industrial competitiveness rather than a stable mechanism for threat mitigation.

Methodology and Analytical Framework
The authors employ a critical policy analysis approach, tracing the historical and intellectual lineage of compute-centric governance.

• Historical Analysis: The paper traces how empirical scaling laws (e.g., power-law relationships between model size, data, and compute) and the "Bitter Lesson" (the efficacy of scaling over hand-crafted features) transformed compute from a research input into a national security lever.
• Policy Critique: The authors analyze specific regulatory instruments, including President Biden's Executive Order 14110 and the Department of Commerce's export controls (2022 and the 2025 "Framework for Artificial Intelligence Diffusion"). They demonstrate how these policies operationalize "frontier" status through input metrics (FLOPs, chip performance) rather than output effects.
• Theoretical Reconstruction: The paper proposes a new regulatory framework grounded in International Humanitarian Law (IHL) and the strategic logic of modern conflict (specifically cost-of-force exchange ratios). It argues for shifting from input-based regulation to effects-based regulation.

Key Contributions

1. Diagnosis of Governance Failure
The paper identifies compute-centric policy as a "proxy-driven governance failure." By treating compute as the object of governance rather than a partial indicator, policy creates false negatives (harmful systems below thresholds) and false positives (high-compute systems with benign uses). This approach concentrates agenda-setting power in the hands of infrastructure incumbents who control the auditable assets, while obscuring risks associated with the "AI Triad" (data, algorithms, compute) when they are decoupled from massive compute budgets.

2. Proposed Definition of AI Weaponization
The authors propose an "intent-and-capability" definition of AI weaponization, independent of computational scale:

• Definition: "The intentional employment or objective immediate capability of an artificial intelligence system to cause physical, digital, or psychological harm, to create an asymmetric advantage in conflict, or to materially facilitate such effects."
• Components:
  • Intent/Capability: Captures both deliberate misuse and systems with immediate objective capability to generate harm (e.g., designing novel pathogens), regardless of developer intent.
  • Harm Domains: Extends beyond kinetic violence to include digital harm (cyber operations) and psychological harm (disinformation/deepfakes).
  • Asymmetric Advantage: Focuses on capabilities that alter wartime outcomes or force-exchange ratios, rather than routine commercial efficiencies.
  • Material Facilitation: Includes systems that significantly reduce the skill, time, or resources required to execute harm, even if they do not directly execute the act.

3. Operationalization via Live Benchmarks
To make this definition actionable, the paper outlines a measurement infrastructure based on live, adversarial benchmarks across the full AI Triad:

• Benchmark Design: Moving beyond static knowledge tests (e.g., "can the model describe a nerve agent?") to procedural competence tests (e.g., "can the model design a synthesis pathway?").
• Triad Evaluation:
  • Hardware: Licensing contingent on testing novel architectures for high-risk capability acceleration.
  • Data: Requiring representative subsets of proprietary datasets for secure benchmarking to assess risk shifts.
  • Algorithms: Continuous evaluation of state-of-the-art algorithms developed on shared resources against high-risk task benchmarks.
• Institutional Architecture: A hybrid model utilizing the National AI Research Resource (NAIRR) for secure, air-gapped evaluation infrastructure and the NIST Center for AI Standards and Innovation (CAISI) for designing validity-tested benchmarks. This structure aims to prevent regulatory capture by industry incumbents.

Results and Evidence
The paper presents evidence that the compute-centric perimeter is already being bypassed:

• Algorithmic Efficiency: Chinese firms have trained state-of-the-art models (e.g., Hunyuan-Large, DeepSeek-R1) on export-compliant GPUs using optimized techniques, demonstrating that static hardware thresholds do not reliably bound capability development.
• Task-Specific Weaponization: Examples include the Ukrainian military deploying small models (millions of parameters) on modest hardware for autonomous targeting, and the creation of specialized datasets (e.g., Zhousidun) to train object detection models for targeting naval assets. These systems operate well below "frontier" compute thresholds but possess direct military utility.
• Policy Volatility: The rapid rescission of the 2025 Diffusion Framework and subsequent approval of high-end chip exports under revenue-sharing arrangements illustrate how the lack of a defined "weaponization" target allows policy to drift toward industrial bargaining.

Significance and Claims
The paper claims that the "LLM Mirage" has held U.S. policy-making hostage by substituting a convenient administrative metric (compute) for the complex reality of weaponization risk.

• Strategic Significance: By focusing on the apex of the AI ecosystem (frontier LLMs), current policy ignores the systems most likely to be weaponized in practice: small, specialized systems that alter force-exchange dynamics by lowering the cost of attack.
• Regulatory Significance: The proposed framework offers a pathway to durable security policy by grounding regulation in demonstrated capability and real-world use cases rather than training inputs. It argues that security policy must align with the effects-based logic of International Humanitarian Law to remain relevant in an era of technological asymmetry.
• Equity Implications: The authors note that compute-centric controls enforce a "security-through-scarcity" model that disproportionately burdens the Global South, enforcing technological underdevelopment while failing to address immediate, task-specific weaponization risks faced by those populations.

The paper concludes that escaping the LLM Mirage requires a shift from inferring risk from training inputs to testing deployable performance, supported by independent, adversarial evaluation infrastructure.