PASA: A Principled Embedding-Space Watermarking Approach for LLM-Generated Text under Semantic-Invariant Attacks

This collection of recent papers highlights new attack vectors, evaluation frameworks, and defenses in LLM- and agent-related security. Key contributions: PASA — a semantic (embedding-space) watermarking method robust to paraphrasing; a low-cost, few-shot “benign” DPO fine-tuning attack that reliably suppresses refusals and enables jailbreaking across multiple OpenAI models; MT-JailBench — a modular benchmark showing prompt generation and evaluation budgets strongly affect multi-turn jailbreak evaluations; a position paper identifying the Authorization‑Execution Gap (AEG) as a structural cause