LinguaJailbreak-Lab: A Crowdsourced Discovery and Analysis Framework for Cross-Lingual Jailbreak Attacks

An open-source research tool based on the CC-BOS method, which uses crowdsourced intelligence to guide the discovery and evaluation of cross-lingual security vulnerabilities in large language models

• Original Author/Maintainer: batis1
• Source Platform: GitHub
• Original Title: LinguaJailbreak-Lab
• Original Link: https://github.com/batis1/LinguaJailbreak-Lab
• Publication Date: May 25, 2026

---

With the global deployment of large language models (LLMs), an often-overlooked security dimension has emerged: cross-lingual attacks. Attackers may exploit the multilingual capabilities of models to bypass safety alignment mechanisms using low-resource languages or classical languages. The LinguaJailbreak-Lab project addresses this challenge by providing a crowdsourced-guided framework for cross-lingual jailbreak attack discovery and analysis

The core of the project is based on the CC-BOS (Classical Chinese-Based Optimization Strategy) method, an optimization strategy that uses classical Chinese as an attack medium. Research shows that classical Chinese, as a semantically rich language with insufficient coverage in modern LLM safety training data, can be an effective attack vector

The project provides complete experimental reproduction configurations:

• Attack Method: CC-BOS
• Attack Language: Classical Chinese
• Target Model: GPT-4o
• Prompt Generation Model: DeepSeek-Chat
• Translation Model: DeepSeek-Chat
• Evaluation Model: GPT-4o
• Population Size: 5
• Maximum Iterations: 5
• Success Criterion: released-code score >= 80
• Early Stop Threshold: score >= 120

The project is designed with two operation modes to meet different research needs:

This mode uses Qwen-Plus to unify the entire process of prompt generation, target response, translation, and evaluation. This design simplifies API management, allowing researchers to verify the complete CC-BOS process with just one API key configuration

This mode strictly follows the implementation of the original CC-BOS paper, using multiple different models for collaborative work, and is suitable for research that requires direct comparison with the paper's results