DeepSeek Harness Plugins · admin-security
Plugins that extend DeepSeek Harness (DSH) with vision, workflows, memory, web tools and more.
A read-only plugin that computes SHA-256 or SHA-512 file checksums through the active Harness filesystem provider without loading file contents into the model conversation.
DSH plugin marketplace with a web UI to manage HTTPS plugin repositories, install or remove plugins, toggle static plugins, and verify SHA-256 checksums.
Structured stack-trace parsing for DSH that recognizes Python, Node.js, Rust, and Java-style errors and extracts bounded frame metadata as compact evidence for the model.
A fail-closed guard that detects secret and token patterns in model-supplied tool arguments before execution.
Gives the model a real Chromium browser for navigation, clicks, screenshots, and tabs, with per-action URL allowlist re-checks, private-network blocking by default, and no model-controlled file paths.
Audits whether a Harness plugin adds value to your current model — pre-install review, post-install audit, and model-switch re-audit reminders.
Adds a password gate, CIDR allowlist, and proxy-header rejection to secure DeepSeek Harness LAN web access.
Adds an LLM-judge auto-approval mode to dsh: low-risk operations pass silently, doubtful ones still prompt, with fail-closed fallback by design.
Deterministic zero-token static security audit for dynamic Cordis plugins in DSH: scan on define, gate on activate, report in-stream, fail-open by design.
Adds password-based authentication to DSH Web GUI covering HTTP, WebSocket, and SSE, with brute-force rate limiting and hot listen-address switching.
Renames the DSH Web UI read-only permission preset to Plan mode with read-only files, enabled web search, and no permission escalation, via a zero-runtime config patch.
In-process data-loss prevention for DeepSeek Harness: denies credential-file access, redacts secrets from tool results and messages, detects payment cards, and strips hidden-instruction characters.