A pentest co-pilot that plugs straight into DeepSeek Harness

What it does

It’s a plugin (JavaScript) for DeepSeek Harness Desktop, activated by the word “富江” (Fujiang). Once installed, it injects a custom system prompt, ships a built-in CVE/PoC knowledge base organized into nine categories (Java deserialization, edge devices, Exchange/AD, cloud/container, Linux privesc, and more), and can download external skill libraries — REA for reverse engineering and AboutSecurity for security assessment — on demand. It also manages a “penetration campaign” ledger that tracks assets, steps, and findings across a staged pipeline from reconnaissance through exploitation to reporting.

The interesting bit

The campaign system is the real novelty: each task binds to one session, and the plugin feeds the model task context via campaign:context so the AI can record findings, advance stages, and get next-step suggestions — with auto-continuation capped at 8 rounds and stall detection. The knowledge base entries carry metadata like verified: false and sources:, so the system at least tries to distinguish “public info” from “confirmed on this target.”

Key highlights

• Local fingerprint matching API (POST /api/dsh-prompt-custom/match) that ranks CVE candidates by port, product, keyword, and severity weights.
• Nine-category knowledge base ships with the plugin — no download needed — and entries are re-read on load so updates take effect without reinstalling.
• Skill libraries (REA, AboutSecurity) are fetched from upstream on demand, not bundled; REA’s npm install disables lifecycle scripts, so Frida’s optional native bindings aren’t built that way.
• Vulnerability reports are persisted locally and browsable by severity (critical/high/medium/low/info).
• Uses only public Harness extension points (systemPrompt, skills, tools, webServer) — no patching of Harness binaries.

Caveats

• The README states plainly that full runtime verification hasn’t been redone — file structure and install logic were checked, but not every skill, CVE entry, or automated flow.
• Known bug: after closing the disclaimer dialog, Enter/Esc listeners can still interfere with input and shortcuts.
• Knowledge entries are curated public material, not verified conclusions; the project itself says exploitation steps need re-checking against target versions.

Verdict

Worth a look if you already run DeepSeek Harness and want a structured, knowledge-backed assistant for authorized pentest work. If you’re not in that ecosystem — or you need verified, production-grade tooling — this is an ambitious hobbyist project with honest disclaimers, not a commercial product.