dsh-plugin-guard

Once you have a handful of third-party plugins installed in the DeepSeek Harness Web GUI, a natural question lingers: what are they actually doing? Are they quietly reading environment variables, opening outbound connections, or forking child processes well beyond their stated purpose? dsh-plugin-guard exists to answer that question without ever executing plugin code. It reads the source and metadata of every installed plugin and renders a green / yellow / red three-tier security health-check report.

The audit runs in three layers. First, a static scan inspects every file against thirteen dangerous patterns (child_process, eval, shell string-building, suspicious external URLs, obfuscation traces, and more), while also flagging non-registry dependencies and preinstall / postinstall scripts as classic supply-chain attack surfaces. Second, a multi-layer reputation check queries npm registry metadata, the OSV.dev vulnerability and malicious-package database, runs bilingual web searches for reported malware or backdoor complaints, and pulls GitHub repository signals such as star count, archive status, and account age. Every lookup is best-effort with automatic graceful degradation, so a missing source never blocks the audit. Third, an AI online audit calls the default model to cross-reference the plugin's self-described function against the static evidence and reputation signals, outputting a safe / suspicious / malicious / inconclusive verdict along with actionable recommendations.

It is well suited for developers and ops engineers who have installed several community plugins and want to confirm none of them are overstepping, as well as small teams looking for a lightweight plugin security review without adopting a full-fledged SAST pipeline. One caveat: this is post-hoc detection, so a rule hit does not prove malice and a clean scan does not prove safety. Final trust decisions should always involve human review.