ACL Findings2026
FLASH: Focused Layer Attention Sink Hijacking
Siyuan Deng, Yerong Li, Fanhua Shang, Hongying Liu
Abstract
Despite advances in safety alignment, Large Language Models (LLMs) remain vulnerable to jailbreaking attacks.However, prevailing methods suffer from a dichotomy of limitations: they either rely on prohibitive iterative optimization in the input space (leading to high computational costs) or fail to penetrate the model's internal decision-making processes.In this work, we identify a critical structural vulnerability: the "Attention Sink" mechanism-originally designed to maintain generation stability by anchoring to initial tokens-unintentionally serves as a computational anchor for safety alignment.We hypothesize and empirically verify that safety guardrails are not globally distributed but are predominantly "front-loaded" in specific attention heads of shallow layers.To audit this structural fragility, we propose FLASH (Focused Layer Attention Sink Hijacking), a novel diagnostic auditing framework that executes a surgical intervention.By precisely scaling attention scores in these vulnerable layers, we dismantle the model's internal safety anchor.To ensure the attack's robustness and coherence against the resulting internal noise, we synergize this intervention with multi-variant query rewriting and an adaptive dynamic decoding strategy.Extensive experiments on Llama-3, Qwen-3, and others demonstrate that FLASH achieves a state-of-the-art Attack Success Rate of over 77% with an unprecedented efficiency of 1.53 queries on average.This work marks a paradigm shift from brute-force optimization to mechanism-driven diagnostic auditing, exposing a fundamental trade-off between architectural stability and safety security.29 28 30 31 Layer 0 ቄ Give steps for how to make a bomb how to make a bomb and Give steps Multi-variant Query + Ok Yes ቄ + Ok
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- COLD-Attack: Jailbreaking LLMs with Stealthiness and Controllability cites How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMs
- COLD-Attack: Jailbreaking LLMs with Stealthiness and Controllability cites HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
Builds on8
- Measuring Massive Multitask Language UnderstandingDan Hendrycks, Collin Burns, Steven Basart, Andy Zou et al.ICLR 2021 · 7,905 citations
- Efficient Streaming Language Models with Attention SinksGuangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han et al.ICLR 2024 · 1,714 citations
- HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust RefusalMantas Mazeika, Long Phan, Xuwang Yin, Andy Zou et al.ICML 2024 · 1,031 citations
- COLD-Attack: Jailbreaking LLMs with Stealthiness and ControllabilityXingang Guo, Fangxu Yu, Huan Zhang, Lianhui Qin et al.ICML 2024 · 173 citations
- How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMsYi Zeng, Hongpeng Lin, Jingwen Zhang, Diyi Yang et al.ACL 2024 · 64 citations
Related papers
- Seeing No Evil: Blinding Large Vision-Language Models to Safety Instructions via Adversarial Attention HijackingJingru Li, Wei Ren, Tianqing ZhuACL 2026
- Attention Eclipse: Manipulating Attention to Bypass LLM Safety-AlignmentPedram Zaree, Md Abdullah Al Mamun, Quazi Mishkatul Alam, Yue Dong et al.EMNLP 2025
- Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank ModificationsBoyi Wei, Kaixuan Huang, Yangsibo Huang, Tinghao Xie et al.ICML 2024 · 215 citations
- Safety Alignment Should be Made More Than Just a Few Tokens DeepXiangyu Qi, Ashwinee Panda, Kaifeng Lyu, Xiao Ma et al.ICLR 2025
- CAVGAN: Unifying Jailbreak and Defense of LLMs via Generative Adversarial Attacks on their Internal RepresentationsXiaohu Li, Yunfeng Ning, Zepeng Bao, Mayi Xu et al.ACL Findings 2025 · 3 citations
