Cover Image

Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism via Probabilistically Ablating Refusal Direction

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Xie, Yuanbo, Zhang, Yingjie, Liu, Tianyun, Ma, Duohe, Liu, Tingwen
Format:	Preprint
Published:	2025
Subjects:	Cryptography and Security
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

SARSteer: Safeguarding Large Audio-Language Models via Safe-Ablated Refusal Steering
by: Lin, Weilin, et al.
Published: (2025)

Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications
by: David, Isaac, et al.
Published: (2026)

FFT: Towards Harmlessness Evaluation and Analysis for LLMs with Factuality, Fairness, Toxicity
by: Cui, Shiyao, et al.
Published: (2023)

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

Self and Cross-Model Distillation for LLMs: Effective Methods for Refusal Pattern Alignment
by: Li, Jie, et al.
Published: (2024)

Why Do Aligned LLMs Remain Jailbreakable: Refusal-Escape Directions, Operator-Level Sources, and Safety-Utility Trade-off
by: Chen, Yu, et al.
Published: (2026)

Defensive Refusal Bias: How Safety Alignment Fails Cyber Defenders
by: Campbell, David, et al.
Published: (2026)

Safety Alignment Should Be Made More Than Just A Few Attention Heads
by: Huang, Chao, et al.
Published: (2025)

LLMs Can Unlearn Refusal with Only 1,000 Benign Samples
by: Guo, Yangyang, et al.
Published: (2026)

Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning?
by: Yin, Qingyu, et al.
Published: (2025)

HumorReject: Decoupling LLM Safety from Refusal Prefix via A Little Humor
by: Wu, Zihui, et al.
Published: (2025)

MCP Safety Training: Learning to Refuse Falsely Benign MCP Exploits using Improved Preference Alignment
by: Halloran, John
Published: (2025)

Safety Context Injection: Inference-Time Safety Alignment via Static Filtering and Agentic Analysis
by: Xu, Zhenhao, et al.
Published: (2026)

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

Safety is Not Only About Refusal: Reasoning-Enhanced Fine-tuning for Interpretable LLM Safety
by: Zhang, Yuyou, et al.
Published: (2025)

Beyond Explicit Refusals: Soft-Failure Attacks on Retrieval-Augmented Generation
by: Zhang, Wentao, et al.
Published: (2026)

Revisiting JBShield: Breaking and Rebuilding Representation-Level Jailbreak Defenses
by: Derya, Kemal, et al.
Published: (2026)

Refusal-Trained LLMs Are Easily Jailbroken As Browser Agents
by: Kumar, Priyanshu, et al.
Published: (2024)

Benign Fine-Tuning Breaks Safety Alignment in Audio LLMs
by: Roh, Jaechul, et al.
Published: (2026)

Beyond Static Pattern Matching? Rethinking Automatic Cryptographic API Misuse Detection in the Era of LLMs
by: Xia, Yifan, et al.
Published: (2024)

Beyond Surface-Level Patterns: An Essence-Driven Defense Framework Against Jailbreak Attacks in LLMs
by: Xiang, Shiyu, et al.
Published: (2025)

SafeThinker: Reasoning about Risk to Deepen Safety Beyond Shallow Alignment
by: Fang, Xianya, et al.
Published: (2026)

Targeting Alignment: Extracting Safety Classifiers of Aligned LLMs
by: Ferrand, Jean-Charles Noirot, et al.
Published: (2025)

Do Reasoning LLMs Refuse What They Infer in Long Contexts?
by: Fu, Yu, et al.
Published: (2026)

Defending MoE LLMs against Harmful Fine-Tuning via Safety Routing Alignment
by: Kim, Jaehan, et al.
Published: (2025)

MTSA: Multi-turn Safety Alignment for LLMs through Multi-round Red-teaming
by: Guo, Weiyang, et al.
Published: (2025)

Agent Safety Alignment via Reinforcement Learning
by: Sha, Zeyang, et al.
Published: (2025)

Refusal Before Decoding: Detecting and Exploiting Refusal Signals in Intermediate LLM Activations
by: Collu, Matteo Gioele, et al.
Published: (2026)

Adversary-Aware DPO: Enhancing Safety Alignment in Vision Language Models via Adversarial Training
by: Weng, Fenghua, et al.
Published: (2025)

GradSafe: Detecting Jailbreak Prompts for LLMs via Safety-Critical Gradient Analysis
by: Xie, Yueqi, et al.
Published: (2024)

Misrouter: Exploiting Routing Mechanisms for Input-Only Attacks on Mixture-of-Experts LLMs
by: Fei, Zekun, et al.
Published: (2026)

Maven-Lockfile: High Integrity Rebuild of Past Java Releases
by: Schmid, Larissa, et al.
Published: (2025)

VisuoAlign: Safety Alignment of LVLMs with Multimodal Tree Search
by: Li, MingSheng, et al.
Published: (2025)

Embedding Poisoning: Bypassing Safety Alignment via Embedding Semantic Shift
by: Yuan, Shuai, et al.
Published: (2025)

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

Jailbreaking LLMs & VLMs: Mechanisms, Evaluation, and Unified Defense
by: Chen, Zejian, et al.
Published: (2026)

RefusalGuard: Geometry-Preserving Fine-Tuning for Safety in LLMs
by: Asif, Sadia, et al.
Published: (2026)

Enhancing Code Vulnerability Detection via Vulnerability-Preserving Data Augmentation
by: Liu, Shangqing, et al.
Published: (2024)

Can LLMs Threaten Human Survival? Benchmarking Potential Existential Threats from LLMs via Prefix Completion
by: Cui, Yu, et al.
Published: (2025)

Immune: Improving Safety Against Jailbreaks in Multi-modal LLMs via Inference-Time Alignment
by: Ghosal, Soumya Suvra, et al.
Published: (2024)