yexi-by/dsh-unrestricted

What it solves

This project provides a plugin for DeepSeek Harness (DSH) that bypasses the model's built-in safety refusals and moralizing disclaimers. It allows users to obtain unrestricted technical outputs for tasks such as software reverse engineering, cybersecurity research, and sensitive creative writing without the model blocking the request or providing "preachy" warnings.

How it works

The plugin hooks into DSH's system prompt assembly pipeline. When enabled, it dynamically injects specialized "jailbreak" instructions directly into the system prompt before the request is sent to the LLM. It is designed to be mode-aware, meaning it adapts its injection strategy to work seamlessly with DSH's various operational modes—such as Plan Mode, Code Execution (PTC), and sub-agent workflows—without breaking the underlying tool-calling protocols or safety sandboxes of the host application.

Who it’s for

It is intended for security researchers, software developers, and creative writers who use DeepSeek Harness and require the model to provide complete, unredacted technical implementations, exploit proofs-of-concept, or specific creative content that standard safety alignment filters might otherwise reject.

Highlights

• Dynamic Injection: Injects rules at the system prompt level, ensuring they take precedence over standard model alignment.
• Mode Compatibility: Fully supports DSH's native features, including Plan Mode, code execution, and multi-agent task delegation.
• Safety-Aware: Does not disable the host application's internal security sandboxes (like file read-only protections or command approval popups).
• Verification System: Includes automated checks to ensure compatibility with DSH updates, automatically disabling itself if the underlying prompt structure changes to prevent errors.
• Non-Destructive: Offers a "no-trace" rollback mechanism that restores original system prompts immediately upon disabling the plugin.