howmp/dsh-pentest

What it solves

It provides a structured penetration testing mode for the DeepSeek Harness (dsh) framework. It allows security researchers to record targets, track exploration clues, verify results, and manage assets and vulnerabilities within a formal workflow, rather than relying on unstructured chat logs.

How it works

The project functions as a plugin for dsh, integrating a specialized storage domain and a set of tools for the AI agent to use. It uses a SQLite backend to store a graph-based domain model consisting of goals, intents, facts, findings, and assets.

Key mechanisms include:

• Tool-based Recording: The agent uses specific tools (e.g., pentest_add_finding, pentest_add_asset) to write data directly to the database.
• Deterministic IDs: Nodes and edges are assigned IDs (like goal-1) allowing the model to reference them across multiple tool calls.
• Session Projection: The system projects the current state of the penetration test into a Web UI, allowing users to visualize the exploration chain, vulnerability lists, and asset maps.
• Protocol Injection: It injects a system prompt (pentest:protocol) that guides the agent to follow the exploration chain and interact in Chinese.

Who it’s for

Security professionals and penetration testers who use DeepSeek Harness to automate or assist in security assessments.

Highlights

• Visual Exploration Chain: Uses @xyflow/react to render a graph showing how goals lead to intents, facts, and findings.
• Structured Asset & Vulnerability Tracking: Dedicated views for managing assets and documenting reproducible steps for vulnerabilities.
• Automated Reporting: Ability to generate Markdown-based test reports from the recorded data.
• Integrated SQLite Backend: Uses a dedicated SQLite database for persistence, separate from the host's default storage.
• Authorization Tracking: Includes a specific field for authorization documentation to ensure audit trails for legal testing.