patronus-protect/patronus-security-cli--plugins-deepseek ↗★ 1
@patronus/deepseek-security
Patronus Runtime Protection and explicit security scans for DeepSeek Harness 适合对AI执行环境安全性要求极高,需要实时防御恶意代码的用户。
インストール
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:patronus-protect/patronus-security-cli#c579b1c1750046dcfbd317a439ea72619ed2118d&path:plugins/deepseekドキュメント
README 全文を読む ↗Patronus Security for DeepSeek Harness
Patronus Runtime Protection integrates with the native Cordis lifecycle in DeepSeek Harness. Each Harness agent gets its own authenticated local scanner session; the integration never bundles or downloads a replacement CLI.
Install
Install the supported DeepSeek Harness CLI first:
npm install --global @deepseek-ai/dsh@0.1.2-rc.1
Then select DeepSeek during Patronus onboarding. It detects dsh, downloads the matching Patronus package, installs it into the headless profile and enables protection automatically:
patronus-security-scanner onboarding
To add DeepSeek later:
patronus-security-scanner integration deepseek install
For installation by an agent, provide INSTALL.md.
Restart the Harness after installation. Try:
Check this repository with Patronus.
Patronus status.
What the native gates do
| Harness event | Patronus behavior |
|---|---|
agent/pre-step | Scans new user-authored text before an agent step begins. |
llm/stream | Enforces the same prompt gate immediately before a model stream as a second native boundary. |
tools/pre-execute | Marks Patronus-owned tools before dispatch so their trusted receipts are not rescanned as external content. |
tools/post-execute | Scans tool-result text and MCP content[].text before the next model turn. |
tools/result | Detects a host projection that changed after the gate and adds degraded context to the next turn. |
agent/disposed | Closes the agent's scanner process and private session handles. |
Multiple text blocks retain their order, and JSON-looking text remains raw. Images, paths, tool names, arguments, envelope keys, metadata and media bytes are outside scanner input.
Runtime flow
- The Harness exposes user input or a completed tool result through its native event bus.
- Patronus selects only the external text covered by the runtime contract.
- The installed CLI scans it under a session capability derived from the real Harness agent identity.
- Clean text continues. Findings return a bounded receipt or an available redacted result without exposing the dangerous original.
- Pending work remains tied to that agent session and can be checked without repeating the source tool.
Exact chat messages patronus off, patronus on and patronus status control protection for the current agent session.
Fail-open behavior
Patronus is fail-open when scanning infrastructure is unavailable. Missing authentication, exhausted API usage, scanner startup errors, timeouts or an unavailable API do not block the underlying Harness tool. The original result continues with an explicit Patronus degraded message and must be treated as unverified. An unavailable scan is never reported as approved.
A completed finding is enforced: dangerous prompt or result text does not continue as an approved original. Fully covered PII/DLP-only results may use the separately retrieved masked view.
Manage the integration
Open patronus-security-scanner dashboard to review activity and change policy. Restart the Harness after installation or an update.
patronus-security-scanner integration deepseek status
patronus-security-scanner integration deepseek update
patronus-security-scanner integration deepseek uninstall
Node.js 22.19 or newer, the pinned dsh release and the matching Patronus Security CLI release are required.
Licensed under Apache-2.0. The package includes LICENSE and THIRD_PARTY_NOTICES.md.