Digests · 6 papers

Paper digests

Each digest works through one paper with PACES: its purpose, approach, claims, evaluation, and what it adds to the taxonomy.

  1. 2026
    I·2 · Tool Manipulation

    Malicious Agent Skills in the Wild: A Large-Scale Security Empirical Study

    Liu, Chen, Zhang, Deng, Li, Ning, et al. · USENIX Security 2026

    The first labelled study of malicious agent skills finds 157 confirmed among 98,380 community skills, most of their attacks written in plain language rather than code, and one actor behind more than half of them.

  2. 2026
    I·3 · Prompt Injection

    ACIArena: Toward Unified Evaluation for Agent Cascading Injection

    An, Li, Zhang, Xu, Zhou, Li, et al. · ACL 2026

    ACIArena benchmarks how one injected input, agent or message cascades through a multi-agent system, across six real frameworks, and finds that topology alone says little about how robust a system is.

  3. 2025
    M·1 · Safety Alignment

    Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs

    Xie, Song, Luo · NeurIPS 2025

    An attacker can jailbreak a model through a fine-tuning API using only ten benign question–answer pairs, by first overfitting it to refuse everything and then teaching it to answer normally.

  4. 2025
    I·2 · RAG & Memory Poisoning

    PoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation of Large Language Models

    Zou, Geng, Wang, Jia · USENIX Security 2025

    PoisonedRAG shows that an attacker who can add a few texts to a retrieval database can make a RAG system give an answer of their choosing, without touching the model or the retriever.

  5. 2025
    I·3 · Prompt Injection

    Red-Teaming LLM Multi-Agent Systems via Communication Attacks

    He, Lin, Dong, Xu, Xing, Liu · ACL 2025 Findings

    Agent-in-the-Middle attacks a multi-agent system without touching any agent, only the messages one agent receives, and succeeds in over 40% of cases in every setting tested.

  6. 2025
    M·1 · Safety Alignment

    Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation

    Huang, Hu, Ilhan, Tekin, Liu · ICLR 2025

    Booster defends against harmful fine-tuning before it happens, by training the aligned model so that a step towards harmful behaviour barely moves it.