dsh-purge is a cybersecurity red-team plugin for DeepSeek Harness that enables controlled testing of model safety policies.
The tool addresses the problem that DeepSeek Harness applies multiple layers of filtering—rewriting prompts, enforcing approval gates, applying sandbox policies, and managing session migration—making it difficult to distinguish between model-level refusals and host-layer blocks. dsh-purge separates these layers on a local installation, allowing security researchers to evaluate which default policies may be overly restrictive and to test model behavior under controlled conditions. It works by patching the host installation with configurable prompts that can be swapped per model, with a default prompt targeting Chinese models.
This tool is designed for security researchers and red-team operators who need to assess the robustness of language model safety mechanisms. It suits projects involving controlled evaluation of model policies and identifying gaps in host-layer protections. The plugin requires DeepSeek Harness to already be installed locally and operates only on that installation. After installation, users must click Apply once, which triggers automatic host restart; subsequent chat sessions will include the injected prompts. The tool requires Node.js version 18 or higher, with version 22 or higher needed on Windows for certain import hook functionality.
The project maintains active development with regular updates and operates as an official plugin within the DeepSeek Harness ecosystem. The tool includes a built-in drill console for testing and provides clear documentation distinguishing its operation from blind prompt rewriting. Development focuses on compatibility with specific host versions, with the current release supporting DeepSeek Harness 0.2 only and explicitly not supporting earlier versions.
