MEWRS proposes Basel-style early warning system for frontier AI risks

Pranav Mehta's new paper draws a direct structural parallel between frontier AI governance today and banking regulation before the 2008 financial crisis. Just as post-2008 reforms like Basel III and Dodd-Frank created macro-prudential oversight to catch systemic risks across institutions, Mehta proposes the Macro-Prudential Early Warning and Response System (MEWRS) for internal AI systems used in labs' research, testing, and production. The core insight: discovering a risk in one model is not the same as acting on it, and reviewing models individually misses the correlated build-up across the entire sector.

MEWRS operates on two complementary layers. Layer A adapts a finder-coordinator-defender model: structured reports on dual-use capabilities, autonomy indicators, and security compromises flow through a government clearinghouse to domain-specific defender working groups that can trigger coordinated responses. Layer B uses three quantitative buffer metrics—Effective Compute-at-Risk (ECAR), Cumulative Red-Team Hours (CRTH), and an Alignment Robustness Score (ARS)—to automatically calibrate operational controls. As capabilities scale faster, safeguards tighten proportionally, much like risk-weighted assets drive capital ratios under Basel. The paper also maps six Basel III mechanisms onto AI governance, identifies seven failure modes with mitigations, and outlines an exercise-based validation plan.