Cover Image

Safety Alignment Should Be Made More Than Just A Few Attention Heads

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Huang, Chao, Zhang, Zefeng, Yue, Juewei, Li, Quangang, Zhang, Chuang, Liu, Tingwen
Format:	Preprint
Published:	2025
Subjects:	Cryptography and Security Artificial Intelligence Computation and Language
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

Safety Alignment Should Be Made More Than Just a Few Tokens Deep
by: Qi, Xiangyu, et al.
Published: (2024)

On the Role of Attention Heads in Large Language Model Safety
by: Zhou, Zhenhong, et al.
Published: (2024)

What Matters For Safety Alignment?
by: Li, Xing, et al.
Published: (2026)

Omni-Safety under Cross-Modality Conflict: Vulnerabilities, Dynamics Mechanisms and Efficient Alignment
by: Wang, Kun, et al.
Published: (2026)

The Sum Leaks More Than Its Parts: Compositional Privacy Risks and Mitigations in Multi-Agent Collaboration
by: Patil, Vaidehi, et al.
Published: (2025)

ReasAlign: Reasoning Enhanced Safety Alignment against Prompt Injection Attack
by: Li, Hao, et al.
Published: (2026)

AgentAlign: Navigating Safety Alignment in the Shift from Informative to Agentic Large Language Models
by: Zhang, Jinchuan, et al.
Published: (2025)

SciSafeEval: A Comprehensive Benchmark for Safety Alignment of Large Language Models in Scientific Tasks
by: Li, Tianhao, et al.
Published: (2024)

Detecting RAG Extraction Attack via Dual-Path Runtime Integrity Game
by: Xie, Yuanbo, et al.
Published: (2026)

Lifelong Safety Alignment for Language Models
by: Wang, Haoyu, et al.
Published: (2025)

The Scales of Justitia: A Comprehensive Survey on Safety Evaluation of LLMs
by: Liu, Songyang, et al.
Published: (2025)

Towards Understanding the Safety Boundaries of DeepSeek Models: Evaluation and Findings
by: Ying, Zonghao, et al.
Published: (2025)

SAGE-RT: Synthetic Alignment data Generation for Safety Evaluation and Red Teaming
by: Kumar, Anurakt, et al.
Published: (2024)

From Threat to Tool: Leveraging Refusal-Aware Injection Attacks for Safety Alignment
by: Chae, Kyubyung, et al.
Published: (2025)

Attention Eclipse: Manipulating Attention to Bypass LLM Safety-Alignment
by: Zaree, Pedram, et al.
Published: (2025)

How Alignment and Jailbreak Work: Explain LLM Safety through Intermediate Hidden States
by: Zhou, Zhenhong, et al.
Published: (2024)

Relevance as a Vulnerability: How Web Retrieval Degrades Safety Alignment in LLM Agents
by: Nawal, Aditya, et al.
Published: (2026)

Mitigating the Safety-utility Trade-off in LLM Alignment via Adaptive Safe Context Learning
by: Wang, Yanbo, et al.
Published: (2026)

Safety Alignment Can Be Not Superficial With Explicit Safety Signals
by: Li, Jianwei, et al.
Published: (2025)

Trojan Activation Attack: Red-Teaming Large Language Models using Activation Steering for Safety-Alignment
by: Wang, Haoran, et al.
Published: (2023)

Decoupled Alignment for Robust Plug-and-Play Adaptation
by: Luo, Haozheng, et al.
Published: (2024)

Depth Charge: Jailbreak Large Language Models from Deep Safety Attention Heads
by: Wu, Jinman, et al.
Published: (2026)

AI Risk Management Should Incorporate Both Safety and Security
by: Qi, Xiangyu, et al.
Published: (2024)

Feint and Attack: Attention-Based Strategies for Jailbreaking and Protecting LLMs
by: Pu, Rui, et al.
Published: (2024)

Agent Tools Orchestration Leaks More: Dataset, Benchmark, and Mitigation
by: Qiao, Yuxuan, et al.
Published: (2025)

Character as a Latent Variable in Large Language Models: A Mechanistic Account of Emergent Misalignment and Conditional Safety Failures
by: Su, Yanghao, et al.
Published: (2026)

Internal Safety Collapse in Frontier Large Language Models
by: Wu, Yutao, et al.
Published: (2026)

ADMIT: Few-shot Knowledge Poisoning Attacks on RAG-based Fact Checking
by: Wu, Yutao, et al.
Published: (2025)

Your Agent, Their Asset: A Real-World Safety Analysis of OpenClaw
by: Wang, Zijun, et al.
Published: (2026)

MMDT: Decoding the Trustworthiness and Safety of Multimodal Foundation Models
by: Xu, Chejian, et al.
Published: (2025)

DataShield: Safety-degrading Data Filtering for LLM Benign Instruction Fine-Tuning
by: Zhang, Junbo, et al.
Published: (2026)

Position: Privacy Is Not Just Memorization!
by: Mireshghallah, Niloofar, et al.
Published: (2025)

FewFedPIT: Towards Privacy-preserving and Few-shot Federated Instruction Tuning
by: Zhang, Zhuo, et al.
Published: (2024)

Improved Few-Shot Jailbreaking Can Circumvent Aligned Language Models and Their Defenses
by: Zheng, Xiaosen, et al.
Published: (2024)

Secure Code Generation via Online Reinforcement Learning with Vulnerability Reward Model
by: Wu, Tianyi, et al.
Published: (2026)

Less is More: Sparse Watermarking in LLMs with Enhanced Text Quality
by: Hoang, Duy C., et al.
Published: (2024)

Reimagining Safety Alignment with An Image
by: Xia, Yifan, et al.
Published: (2025)

SafeKey: Amplifying Aha-Moment Insights for Safety Reasoning
by: Zhou, Kaiwen, et al.
Published: (2025)

Geneshift: Impact of different scenario shift on Jailbreaking LLM
by: Wu, Tianyi, et al.
Published: (2025)

LLM Agents Should Employ Security Principles
by: Zhang, Kaiyuan, et al.
Published: (2025)