Behavioral Geometry Predicts and Mitigates Jailbreak Susceptibility Across AI Models

A new framework formalizes the behavioral geometry of model populations to predict jailbreak susceptibility and transfer defenses efficiently. Applied to 79 models from 24 providers and 100 configurations of a single base model, simple methods achieve an AUPRC of 0.94 for susceptibility detection using approximately 98% fewer probes than full evaluation. Transferring optimized defenses via behavioral geometry outperforms same-provider assignment by 2% (p=0.03) at no additional probe cost. The approach leverages previously evaluated and defended models to avoid impractical per-configuration evaluation.