dsh-purge: Testing DeepSeek Harness Policy Boundaries in a Local Lab
YuJunZhiXue/dsh-purge
About the project
A cybersecurity red team plugin for analyzing the runtime behavior of DeepSeek Harness. It distinguishes between model refusals and host-layer blocks, which occur at different layers, enabling the evaluation of whether specific policies are overly restrictive. It supports both the official Web and desktop EXE environments and targets DSH 0.2.1-rc.2.
Accessed via a dock button next to the session title, it consists of two pages: Clean and Drill. The Clean page handles patch status checks, application and recovery, prompt editing, and multi-rule set management. The Drill page handles local environments and assets after authorization; this build does not include a skill library or knowledge base.
It is designed to block public internet access, allowing only loopback and RFC1918 addresses. Attempts to resolve domain names to public IPs or connect to public networks via hosts file manipulation are blocked. Misuse attempts, such as game cheat generation or memory editing, are blocked simultaneously at the prompt and session gate levels.
It operates only in a local official Harness environment where the user has administrative privileges and does not target other people's systems or public servers. It is distributed under the Apache-2.0 license, and the original repository source must be specified when forking. Use for illegal purposes is strictly prohibited, and all legal liability lies with the user.
YuJunZhiXue/dsh-purge
DeepSeek Harness 破甲:让所有模型都能破甲,不同模型可换不同提示词;默认提示词面向国模「小码酱」。Jailbreak for every model — swap prompts per model. 求 Star 收藏 ⭐ 如要二开,请查看署名要求部分
JavaScript
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.