Dual-use Risks of Frontier Artificial Intelligence

<p><b>Background. </b>The dual-use risks of frontier artificial intelligence systems have moved from prospective concern to operational reality. Anthropic's activation of AI Safety Level 3 safeguards for Claude Opus 4 in May 2025, explicitly because chemical, biological, radiological, and nuclear (CBRN) uplift risks could no longer be confidently ruled out, marked a watershed in the industry's acknowledgement of risk. The United Kingdom AI Security Institute's first public Frontier Trends Report documented frontier model performance on apprentice-level cyber tasks rising from sub-nine per cent in late 2023 to approximately fifty per cent by 2026, with the first model completing expert-level tasks in 2025. The regulatory architecture intended to manage these risks remains unsettled: the United States Bureau of Industry and Security issued a Framework for Artificial Intelligence Diffusion in January 2025 and rescinded it in May 2025, leaving a regulatory gap that subsequent guidance has only partially filled.</p> <p><b>Purpose. </b>This paper provides a synthesis of dual-use risk evidence for frontier AI across cyber offensive uplift and CBRN uplift, maps the contemporary regulatory architecture including its current volatility, examines the diversion and weight-exfiltration risks that connect dual-use concerns to AI supply chain governance, and analyses the Gulf Cooperation Council strategic context where the United Arab Emirates and the Kingdom of Saudi Arabia occupy a distinctive position in global AI diffusion policy.</p> <p><b>Approach. </b>The paper adopts a narrative review methodology combining contemporary 2024 to 2026 empirical evaluations of frontier AI capabilities (including the UK AI Security Institute trends report, Anthropic Frontier Red Team disclosures, Frontier Model Forum threshold work, and independent third-party assessments), primary regulatory document analysis (the BIS Framework for AI Diffusion and its rescission, the EU AI Act systemic risk provisions, EU Regulation 2021 slash 821 on dual-use exports, the UAE Strategic Goods regime under Cabinet Resolution 50 of 2020), and analysis of voluntary industry safety frameworks (the Anthropic Responsible Scaling Policy, Google DeepMind Frontier Safety Framework version 2.0, OpenAI Preparedness Framework, and parallel instruments from Microsoft, Amazon, and G42).</p> <p><b>Findings. </b>Three findings emerge. First, the empirical evidence base for cyber offensive uplift is now substantial, with multiple converging signals indicating that frontier models provide meaningful tactical uplift to human cyber operators, particularly when paired with appropriate scaffolding; the evidence base for CBRN uplift is more contested due to sensitivity restrictions on full evaluation data, but is sufficient to support the precautionary safeguards that leading labs have implemented. Second, the regulatory architecture is in active flux: the AI Diffusion Rule episode illustrates that unilateral commodity-based export controls struggle with a technology that diffuses through training compute, weights distribution, and inference access in ways that traditional dual-use frameworks were not designed to capture. Third, the GCC strategic context is distinctive: the United Arab Emirates and Saudi Arabia are positioned as Tier 2 destinations in US export-control thinking, with the UAE in particular having demonstrated a willingness to accommodate US security concerns through G42's divestment of Chinese hardware and the March 2025 US-UAE 1.4 trillion technology investment framework.</p> <p><b>Implications. </b>Practitioners deploying frontier AI in the GCC region operate under a multi-layered governance regime that combines formal export controls, voluntary industry safety frameworks, and emerging EU AI Act systemic risk obligations. The diversion risks examined in Section 5 connect directly to the AI Supply Chain Trust Boundary Model developed in Paper 10 of this slate, with weight exfiltration at Layer 2 representing the operational manifestation of the most consequential dual-use diversion concern. Future work will extend the analysis through comparative regulatory mapping (Paper 14) and operational identity considerations, including weight-level attestation (Paper 15).</p>