YuJunZhiXue/dsh-purge

What it solves

dsh-purge is a cybersecurity red-teaming tool designed for the official DeepSeek Harness. It allows researchers to separate the host-layer blocks (such as prompt rewriting, approval gates, and sandbox policies) from the model's actual refusals. This helps in evaluating whether a model is refusing a prompt or if the host software is blocking it, enabling a more controlled robustness evaluation.

How it works

The tool operates as a plugin for the DeepSeek Harness (supporting both the official web and desktop versions). It does not patch the source tree but instead modifies local configuration and packages on the user's machine. It provides a "Clean" page for applying and restoring patches and a "Drill" page for authorized environment assessment. The plugin uses encrypted prompts to inject specific identities or instructions into the session after the user applies the patches and restarts the application.

Who it’s for

It is intended for cybersecurity researchers and red-teamers who have the official DeepSeek Harness installed locally and wish to perform authorized robustness evaluations and security research.

Highlights

• Host-Layer Separation: Distinguishes between host-level policy blocks and model-level refusals.
• Dual Host Support: Compatible with official dsh web and the official desktop EXE.
• Integrated Management: Includes a "Clean" dock for patch status, prompt editing, and rule set management.
• Safety Controls: Specifically designed to run only on the user's own local official Harness installation, not as a public scanner or attack kit.