Minglink/dsh-infinite-gen-4

What it solves

It provides a specialized red-teaming and robustness benchmarking suite for DeepSeek LLMs (specifically V4.1, V4-Pro, and Flash). The project helps security researchers identify defense blind spots, such as prompt injection and unexpected model behaviors, to improve safety guardrails and alignment.

How it works

The tool functions as a plugin for the DeepSeek Harness environment. It uses a "Dual-Layer Kernel" architecture to inject specific system prompts (a general kernel and a reinforced mirror kernel) into the model. It employs a "zero-tool surface" design to eliminate decision noise during the model's thinking phase, forcing direct output through specific token induction and strict output contracts.

Who it’s for

It is designed for cybersecurity red-teamers, academic researchers in AI safety, and security engineering teams who need to conduct controlled robustness testing in offline or authorized laboratory environments.

Highlights

• Dual-Layer Injection: Uses a coordinated system of general and reinforced kernels for more effective prompt injection.
• Zero-Tool Surface: Removes tool-calling noise to ensure faster, direct responses.
• Automated Benchmarking: Includes a regression test suite with over 100 assertions to verify prompt completeness and model behavior.
• Easy Installation: Supports one-click installation via the dsh:// protocol for desktop clients or local scripts for Windows, Linux, and macOS.