Introduction

DeepSeek Harness (DSH) adopts an “everything is plugins” architecture. When adopting third-party plugins, developers often find it difficult to evaluate their potential risks before execution or cannot intuitively determine whether a plugin exceeds its authorized scope. The dsh-plugin-guard plugin performs a security health check on installed plugins through static code analysis, dependency review, and AI online auditing, presenting results in a green/yellow/red three-level panel.

What Is This

This is a plugin security health check tool for DeepSeek Harness Web GUI. Without executing plugin code, it reads the source code and metadata of installed third-party plugins to help determine their functional scope, whether they exhibit unauthorized behavior, and their risk level, and finally generates a tiered report.

Core Features

Static Scanning and Dependency Review

The plugin statically scans installed plugin source code, identifying 17 categories of risky capabilities, including dangerous APIs (such as subprocess, eval, vm, shell), secret leakage, high-entropy payloads, and download-then-execute behavior. It also reviews dependencies and flags non-npm sources (such as git:, file:, URL) and dependencies matching suspicious keywords (such as miner, stealer, backdoor).

Install Script and Permission Scoring

The plugin separately reviews preinstall, install, and postinstall scripts, paying special attention to download-then-execute behavior. It also scores the host service permissions declared by plugins based on entry.inject in dsh.plugin.json and dsh.client.inject in package.json, and detects “capability and declared surface mismatches” (for example, code uses high-risk capabilities, but declared permissions are lightweight).

Version Monitoring and Alerts

The plugin stores baseline data and continuously monitors installed plugins. If a new installation, version change, or new risky capability/declared permission compared to the previous scan is detected, the panel marks it with a badge.

AI Online Auditing and Tiering

The plugin combines code, version, and multi-layer internet reputation (npm metadata, OSV.dev, malware reports, GitHub signals) for secondary AI judgment. After AI auditing, it reclassifies results: malicious is red, suspicious is yellow, safe is green, and inconclusive is no lower than the static level. When AI and static results disagree, the panel explicitly annotates it.

Reputation Evidence

Reputation evidence includes npm registry metadata (package age, deprecation flags), authoritative OSV.dev records, internet malware/attack report searches, and GitHub repository signals. A GitHub Token can increase API rate limits.

Skill Auditing (SafeSkill)

The plugin integrates with the ThreatBook SafeSkill platform, supporting multi-engine scanning (LLM / static / dynamic) of installed DSH Skills. When the SafeSkill quota is exhausted, local AI review can be performed using the default model.

Installation and Enablement

Prerequisites

  1. DSH is installed on the machine and can run dsh web normally.
  2. pnpm is installed on the machine (dsh plugin internally requires it to install plugins).

Installation

Run the following command to install the plugin:

dsh plugin --profile <name> add @guojin-ai/dsh-plugin-guard

Replace <name> with the name of the profile you want to audit. After installation is complete, restart dsh.

Enable and View

After restarting dsh, navigate to Settings → Plugin Security Health Check in the Web GUI to view the security report.

Usage Guide

View Reports

The top of the panel displays summary statistics. Each row represents an installed plugin and includes a risk level badge, plugin name, version, risk score, matched rules, declared permissions, and so on. Expanding a row shows changes since the last scan, matched risk rules, declared host service permissions, and suspicious dependencies.

AI Auditing

Click the “AI Audit” button in a plugin row to view the judgment conclusion (safe/suspicious/malicious/inconclusive) and recommendations. Audit results are cached (content fingerprint + version + TTL), and new negative signals force re-auditing.

SafeSkill Scanning

After configuring the SafeSkill API Key at the top of the panel, click “Scan All Skills”. The plugin packages installed DSH Skills and uploads them to the SafeSkill platform, then retrieves the multi-engine scanning report.

Notes

  1. Package Name Matching: The name in the DSH loading entry must exactly match @guojin-ai/dsh-plugin-guard in package.json; otherwise, the client bundle will be silently discarded.
  2. Data Storage: GitHub Token and SafeSkill Key are stored only locally and are not uploaded.
  3. Platform Independence: SafeSkill is an independent external platform, and scan results are provided by ThreatBook.

Conclusion

dsh-plugin-guard provides a visual security assessment mechanism for DSH plugin usage through static analysis, dependency review, and AI auditing. It is suitable for use before adopting third-party plugins or during periodic inspections, helping developers identify potential supply chain risks.

Plugin Catalog
GitHub Repository