ShallowJail: Steering Jailbreaks against Large Language Models

2026-02-06

AI Summary

Aligned large language models remain vulnerable to jailbreak attacks, yet existing approaches often rely on explicit prompts or incur high computational overhead, struggling to balance stealth and efficiency. This work proposes ShallowJail, which uncovers and exploits the fragility of shallow alignment mechanisms in large language models. By manipulating the initial tokens of model inputs during inference, ShallowJail enables low-overhead, highly stealthy white-box and gray-box jailbreak attacks without requiring complex optimization. Experiments across multiple state-of-the-art large language models demonstrate that ShallowJail significantly compromises their safety guarantees, thereby validating its effectiveness and generalizability.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Large Language Models(LLMs) have been successful in numerous fields. Alignment has usually been applied to prevent them from harmful purposes. However, aligned LLMs remain vulnerable to jailbreak attacks that deliberately mislead them into producing harmful outputs. Existing jailbreaks are either black-box, using carefully crafted, unstealthy prompts, or white-box, requiring resource-intensive computation. In light of these challenges, we introduce ShallowJail, a novel attack that exploits shallow alignment in LLMs. ShallowJail can misguide LLMs'responses by manipulating the initial tokens during inference. Through extensive experiments, we demonstrate the effectiveness of~\shallow, which substantially degrades the safety of state-of-the-art LLM responses.

Problem

• jailbreak attacks
• Large Language Models
• alignment
• harmful outputs
• model safety

Innovation

• jailbreak attack
• shallow alignment
• large language models
• token manipulation
• model safety