Imagen de Portada

Why Do Aligned LLMs Remain Jailbreakable: Refusal-Escape Directions, Operator-Level Sources, and Safety-Utility Trade-off

Fuente: arXiv

Guardado en:

Detalles Bibliográficos
Autores principales:	Chen, Yu, Liu, Yuanhao, Cao, Qi
Formato:	Preprint
Publicado:	2026
Materias:	Cryptography and Security Artificial Intelligence
Acceso en línea:	Acceder al recurso
Etiquetas:	Agregar Etiqueta Sin Etiquetas, Sea el primero en etiquetar este registro!

Ejemplares similares

Jailbreaking Leading Safety-Aligned LLMs with Simple Adaptive Attacks
por: Andriushchenko, Maksym, et al.
Publicado: (2024)

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

Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism via Probabilistically Ablating Refusal Direction
por: Xie, Yuanbo, et al.
Publicado: (2025)

Tracing the Dynamics of Refusal: Exploiting Latent Refusal Trajectories for Robust Jailbreak Detection
por: Hu, Xulin, et al.
Publicado: (2026)

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

JailPO: A Novel Black-box Jailbreak Framework via Preference Optimization against Aligned LLMs
por: Li, Hongyi, et al.
Publicado: (2024)

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

Activation Approximations Can Incur Safety Vulnerabilities Even in Aligned LLMs: Comprehensive Analysis and Defense
por: Zhang, Jiawen, et al.
Publicado: (2025)

The Privacy-Utility Trade-off in the Topics API
por: Alvim, Mário S., et al.
Publicado: (2024)

FlipAttack: Jailbreak LLMs via Flipping
por: Liu, Yue, et al.
Publicado: (2024)

Babel: Jailbreaking Safety Attention via Obfuscation Distribution Optimized Sampling
por: Wang, Ziwei, et al.
Publicado: (2026)

Exploring Jailbreak Attacks on LLMs through Intent Concealment and Diversion
por: Cui, Tiehan, et al.
Publicado: (2025)

PAPILLON: Efficient and Stealthy Fuzz Testing-Powered Jailbreaks for LLMs
por: Gong, Xueluan, et al.
Publicado: (2024)

Gradient Cuff: Detecting Jailbreak Attacks on Large Language Models by Exploring Refusal Loss Landscapes
por: Hu, Xiaomeng, et al.
Publicado: (2024)

A Unified Learn-to-Distort-Data Framework for Privacy-Utility Trade-off in Trustworthy Federated Learning
por: Zhang, Xiaojin, et al.
Publicado: (2024)

Align is not Enough: Multimodal Universal Jailbreak Attack against Multimodal Large Language Models
por: Wang, Youze, et al.
Publicado: (2025)

Clients Collaborate: Flexible Differentially Private Federated Learning with Guaranteed Improvement of Utility-Privacy Trade-off
por: Li, Yuecheng, et al.
Publicado: (2024)

Evaluating Jailbreaking Vulnerabilities in LLMs Deployed as Assistants for Smart Grid Operations: A Benchmark Against NERC Standards
por: Hammadia, Taha, et al.
Publicado: (2026)

Competition Report: Finding Universal Jailbreak Backdoors in Aligned LLMs
por: Rando, Javier, et al.
Publicado: (2024)

ForgeDAN: An Evolutionary Framework for Jailbreaking Aligned Large Language Models
por: Cheng, Siyang, et al.
Publicado: (2025)

Unlearning with Asymmetric Sources: Improved Unlearning-Utility Trade-off with Public Data
por: Inane, Ahmed Mehdi, et al.
Publicado: (2026)

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

Confusion is the Final Barrier: Rethinking Jailbreak Evaluation and Investigating the Real Misuse Threat of LLMs
por: Yan, Yu, et al.
Publicado: (2025)

Enhancing Model Defense Against Jailbreaks with Proactive Safety Reasoning
por: Yang, Xianglin, et al.
Publicado: (2025)

Data to Defense: The Role of Curation in Customizing LLMs Against Jailbreaking Attacks
por: Liu, Xiaoqun, et al.
Publicado: (2024)

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

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

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

Analysis of LLMs Against Prompt Injection and Jailbreak Attacks
por: Jaiswal, Piyush, et al.
Publicado: (2026)

Re-Triggering Safeguards within LLMs for Jailbreak Detection
por: Lin, Zheng, et al.
Publicado: (2026)

PUZZLED: Jailbreaking LLMs through Word-Based Puzzles
por: Ahn, Yelim, et al.
Publicado: (2025)

Enhancing Jailbreak Attacks on LLMs via Persona Prompts
por: Zhang, Zheng, et al.
Publicado: (2025)

LLMs Caught in the Crossfire: Malware Requests and Jailbreak Challenges
por: Li, Haoyang, et al.
Publicado: (2025)

Bag of Tricks: Benchmarking of Jailbreak Attacks on LLMs
por: Xu, Zhao, et al.
Publicado: (2024)

Do Not Merge My Model! Safeguarding Open-Source LLMs Against Unauthorized Model Merging
por: Li, Qinfeng, et al.
Publicado: (2025)

Revisiting Privacy-Utility Trade-off for DP Training with Pre-existing Knowledge
por: Zheng, Yu, et al.
Publicado: (2024)

"To Survive, I Must Defect": Jailbreaking LLMs via the Game-Theory Scenarios
por: Sun, Zhen, et al.
Publicado: (2025)

One Leak Away: How Pretrained Model Exposure Amplifies Jailbreak Risks in Finetuned LLMs
por: Tan, Yixin, et al.
Publicado: (2025)

SafeDream: Safety World Model for Proactive Early Jailbreak Detection
por: Yan, Bo, et al.
Publicado: (2026)

LLMs Can Defend Themselves Against Jailbreaking in a Practical Manner: A Vision Paper
por: Wu, Daoyuan, et al.
Publicado: (2024)