Overview

DeepSeek Harness 破甲 (dsh-purge) is a powerful network security red teaming tool designed to enhance the capabilities of every Large Language Model (LLM) integrated with DeepSeek Harness. This tool allows for flexible prompt customization per model, with default prompts tailored for Chinese domestic models like "小码酱." It functions as a plugin for the official DeepSeek Harness Web and desktop applications, enabling advanced prompt injection and policy modification for security testing and evaluation.

Key Features

• Universal Jailbreaking: Empowers any LLM within DeepSeek Harness to bypass restrictions by swapping prompts tailored to individual models.
• Customizable Prompts: Supports swapping prompts per model, allowing for highly specific and effective red teaming scenarios.
• Deep Integration: Works seamlessly with both the official DeepSeek Harness Web (dsh web) and desktop (EXE) applications, requiring separate installations and applications for each.
• Red Team Focused: Provides tools for network security red team assessments, including prompt manipulation, policy adjustments, and sandbox enhancements.
• Controlled Testing Environment: Separates prompt injection from host interception, enabling controlled evaluations and the identification of potential security vulnerabilities in default policies.
• User-Friendly Interface: Features a dedicated side panel with "Cleanse" and "Drill Ground" sections for managing patches, editing prompts, and accessing asset information (post-authorization).
• Non-Intrusive Patching: Modifies local official packages and configurations without altering the Harness source code, applying patches via the "Apply" function within the plugin.
• Secure and Local: Operates solely on the user's local machine and official Harness installation, acting as a research and evaluation tool, not an external attack suite.

Typical Use Cases

• LLM Security Auditing: Evaluate the robustness of LLMs against various adversarial prompts and jailbreaking attempts.
• Red Team Operations: Enhance red team capabilities by tailoring prompts and modifying model behavior for penetration testing and security assessments within authorized environments.
• Policy Evaluation: Analyze and refine default security policies by observing how different models respond to specific prompt injections.
• Custom Model Integration: Develop and test bespoke prompt strategies for specialized LLMs or domestic Chinese models.
• Security Research: Facilitate research into LLM vulnerabilities and defense mechanisms in a controlled, local setting.