Advanced AI systems bring significant benefits but also introduce risks tied to how their capabilities spread. While replication and incremental research are well understood, less attention has been paid to alternative diffusion mechanisms like theft, espionage, leaks, or extortion. These could accelerate unsafe proliferation or concentrate power in malicious hands. Understanding these mechanisms—their historical precedents, incentives, and possible mitigations—could help shape policies to manage AI risks more effectively.

Exploring Understudied AI Diffusion Risks

One way to address this gap is by systematically investigating four key diffusion mechanisms:

• Leaks: Unintentional disclosures, such as model weights being posted online.
• Theft: Unauthorized access to proprietary AI systems, like hacking into a research labli>
• Espionage: State-sponsored acquisition of AI secrets, such as infiltrating a research team.
• Extortion: Coercing access, like ransomware attacks targeting AI infrastructure.

For each mechanism, research could map incentives (e.g., cost savings, competitive advantage), analyze historical parallels (e.g., nuclear espionage during the Cold War), and propose targeted interventions (e.g., secure model-weight distribution protocols).

Stakeholders and Execution

This research could benefit:

• Policymakers: By providing evidence-based strategies to regulate AI diffusion risks.
• AI Labs: By offering threat models to secure systems against theft or leaks.
• Cybersecurity Experts: By adapting existing tools to AI-specific risks.

An execution plan might involve:

1. Phase 1: Literature review, expert interviews, and incentive modeling to compare mechanisms.
2. Phase 2: Synthesizing findings into a framework ranking risks by severity and tractability, followed by workshops to test interventions.

Differentiating from Existing Work

While some organizations study AI's geopolitical impacts or cybersecurity risks, this approach would focus specifically on AI diffusion mechanisms. For example, it could adapt frameworks from nuclear security research to digital assets like AI models, or tailor cybersecurity insights to AI's unique risks (e.g., model exfiltration). The goal would be to provide granular, actionable recommendations rather than broad analyses.

By addressing these understudied risks, this research could help shape policies and security practices to prevent harmful AI proliferation.