# kl-divergence

DPO-Specific Attacks

Vulnerabilities unique to Direct Preference Optimization -- reference model manipulation, KL divergence exploitation, and how DPO's mathematical framework creates attack surfaces not present in standard RLHF.