New taxonomy exposes gaps in enforcing international AI agreements

A new paper from researchers Raymond Koopmanschap and Otto Barten, accepted at the International Conference on Large-Scale AI Risks 2026, introduces a structured taxonomy for evaluating enforcement mechanisms in international AI agreements. The core principle: wherever sufficient compute capacity exists to violate an agreement, that capacity must be under a control regime. This breaks the enforcement problem into three sub-problems: preventing uncontrolled resource acquisition (e.g., clandestine GPU farms), detecting all compute capacity outside the control regime, and preventing escape from the regime itself.

The authors analyze existing policy proposals through this lens and find a significant imbalance: most efforts target preventing escape (e.g., hardware locks, monitoring of data centers), while the equally critical problems of detecting hidden compute and stopping resource acquisition receive far less attention. They also define an 'enforcement breaking point' — the FLOP threshold (or equivalent metric) at which a policy loses effectiveness. As the compute required for dangerous AI capabilities decreases over time, more actors can violate agreements, making enforcement progressively harder. The taxonomy provides a systematic way for policymakers to identify where enforceability breaks down first and prioritize efforts accordingly.