Attention Slipping: A Mechanistic Understanding of Jailbreak Attacks and Defenses in LLMs

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Hu, Xiaomeng, Chen, Pin-Yu, Ho, Tsung-Yi
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912468572307456
author Hu, Xiaomeng
Chen, Pin-Yu
Ho, Tsung-Yi
author_facet Hu, Xiaomeng
Chen, Pin-Yu
Ho, Tsung-Yi
contents As large language models (LLMs) become more integral to society and technology, ensuring their safety becomes essential. Jailbreak attacks exploit vulnerabilities to bypass safety guardrails, posing a significant threat. However, the mechanisms enabling these attacks are not well understood. In this paper, we reveal a universal phenomenon that occurs during jailbreak attacks: Attention Slipping. During this phenomenon, the model gradually reduces the attention it allocates to unsafe requests in a user query during the attack process, ultimately causing a jailbreak. We show Attention Slipping is consistent across various jailbreak methods, including gradient-based token replacement, prompt-level template refinement, and in-context learning. Additionally, we evaluate two defenses based on query perturbation, Token Highlighter and SmoothLLM, and find they indirectly mitigate Attention Slipping, with their effectiveness positively correlated with the degree of mitigation achieved. Inspired by this finding, we propose Attention Sharpening, a new defense that directly counters Attention Slipping by sharpening the attention score distribution using temperature scaling. Experiments on four leading LLMs (Gemma2-9B-It, Llama3.1-8B-It, Qwen2.5-7B-It, Mistral-7B-It v0.2) show that our method effectively resists various jailbreak attacks while maintaining performance on benign tasks on AlpacaEval. Importantly, Attention Sharpening introduces no additional computational or memory overhead, making it an efficient and practical solution for real-world deployment.
format Preprint
id arxiv_https___arxiv_org_abs_2507_04365
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Attention Slipping: A Mechanistic Understanding of Jailbreak Attacks and Defenses in LLMs
Hu, Xiaomeng
Chen, Pin-Yu
Ho, Tsung-Yi
Cryptography and Security
Artificial Intelligence
Computation and Language
As large language models (LLMs) become more integral to society and technology, ensuring their safety becomes essential. Jailbreak attacks exploit vulnerabilities to bypass safety guardrails, posing a significant threat. However, the mechanisms enabling these attacks are not well understood. In this paper, we reveal a universal phenomenon that occurs during jailbreak attacks: Attention Slipping. During this phenomenon, the model gradually reduces the attention it allocates to unsafe requests in a user query during the attack process, ultimately causing a jailbreak. We show Attention Slipping is consistent across various jailbreak methods, including gradient-based token replacement, prompt-level template refinement, and in-context learning. Additionally, we evaluate two defenses based on query perturbation, Token Highlighter and SmoothLLM, and find they indirectly mitigate Attention Slipping, with their effectiveness positively correlated with the degree of mitigation achieved. Inspired by this finding, we propose Attention Sharpening, a new defense that directly counters Attention Slipping by sharpening the attention score distribution using temperature scaling. Experiments on four leading LLMs (Gemma2-9B-It, Llama3.1-8B-It, Qwen2.5-7B-It, Mistral-7B-It v0.2) show that our method effectively resists various jailbreak attacks while maintaining performance on benign tasks on AlpacaEval. Importantly, Attention Sharpening introduces no additional computational or memory overhead, making it an efficient and practical solution for real-world deployment.
title Attention Slipping: A Mechanistic Understanding of Jailbreak Attacks and Defenses in LLMs
topic Cryptography and Security
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2507.04365