dsh-plugin-trustlens

Before enabling or updating a third-party plugin in DeepSeek Harness, most operators have no systematic security gate. A malicious bundle can carry remote-execution payloads, environment exfiltration, or encoded-evaluation tricks that are easy to miss in manual code review. TrustLens slots a structured security review into the enable pipeline so that administrators see a concrete risk report before the plugin goes live.

The review works in three layers. A read-only static scanner flags common dangerous patterns such as remote execution, download-and-run, encoded eval, persistence mechanisms, and environment exfiltration. A semantic pass runs through whichever model the current DSH session has selected, with no hard-coded backend, keeping the check consistent with session context. A third check reports contradictions between code comments and the README, surfacing cases where documentation and actual behaviour diverge. Throughout, the auditor never imports, executes, or downloads the plugin under inspection; all code, comments, strings, and model output are treated as untrusted data. When the static scan surfaces a high-risk finding, the enable and update actions are locked, and the user can only choose isolation or disable after reviewing the full report.

TrustLens is aimed at DevOps or security leads who manage a growing set of third-party plugins in a DSH environment, especially where plugin origins are diverse, update cadence is high, and an audit trail is required. It does not replace a sandbox or runtime protection layer; it is the last gate before activation, catching the most obvious risks at minimal cost.