MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models

This feed collects multiple 2026 ML-security research contributions that raise urgent integrity, privacy, and supply-chain risks for deployed LLMs and agents. Key findings: MaskForge — a fully black-box, structure-aware attack — achieves ~79% average jailbreak success vs. diffusion LLMs, exposing infill-native exploit surfaces; Covert Influence demonstrates covert, human-undetectable behavioral payload transfer between models across fine-tuning, distillation, and in-context learning; credential-exfiltration work shows feasible pre-output and multi-turn attacks and proposes activation-based,_hh