dsh-plugin-safety

When you install a third-party plugin in DeepSeek Harness, the UI typically offers a single "Confirm install" button while the package may already contain remote code execution chains, credential-harvesting lifecycle scripts, or system persistence mechanisms. dsh-plugin-safety closes that gap: every install request routed through the plugin market (dsh-market or Community Market) is intercepted at the shared fetch layer and subjected to a full static scan plus metadata analysis on the host side before anything is allowed through, using nothing but Node built-in modules.

The scan inspects JavaScript and TypeScript source inside npm tarballs and GitHub repository archives for a dozen-plus signal patterns, including download-eval-spawn triads, dynamic code construction (eval, new Function, vm), suspicious postinstall scripts that curl remote payloads, hard-coded public IPs and cloud-metadata endpoints, persistence via launchctl / Windows Registry / systemd, credential-file paths appearing alongside file I/O within a narrow window, and large Base64 or decryption-routine obfuscation. Metadata from npm and GitHub such as days since publish, download volume, missing repository or license, and look-alike package names detected via Levenshtein distance are layered on top with weighted scoring. Every finding carries a rule name, severity level, and an evidence snippet, so the resulting caution or risky verdict is fully explainable rather than a black-box judgment. Concurrent installs share a single cached scan (six-hour TTL), and any code that could not be read escalates the result to a caution prompt instead of silently passing.

With zero runtime dependencies the tool keeps its own attack surface minimal. It is built for DeepSeek Harness users who regularly install extensions from the marketplace or Community Market and want a meaningful safety net. It will not catch a perfectly obfuscated zero-day payload, but it intercepts the patterns behind the vast majority of malicious or negligent plugins and tells you, item by item, exactly why each one was flagged.