OpenTrustBench - Trust & security scanner for AI agents and MCP ser

项目说明

<p align="center">
<a href="
<img src=" alt="OpenTrustBench Grade A — opentrustbench CLI self-scan" width="120">
</a>
<br>
_{Self-scan of <code>packages/cli</code> at HEAD: <strong>A (90/100)</strong>, 0 findings, minimal scope. Reproduce: <code>node packages/cli/dist/index.js scan packages/cli --quiet</code>}
<br>
<strong>Trust &amp; Security for AI Agents &amp; MCP Servers</strong>
</p>

<h1 align="center">OpenTrustBench</h1>

<p align="center">
<a href=" src=" alt="npm "></a>
<a href=" src=" alt="License"></a>
<a href=" src=" alt="OWASP Agentic AI"></a>
<a href=" src=" alt="8 OWASP-mapped rules"></a>
</p>

<p align="center">
Scan your AI agent or MCP server for .<br>
Get a <strong>Trust Card</strong> with a public grade badge.
</p>

---

What is OpenTrustBench?

OpenTrustBench is a trust and security platform for autonomous AI agents and MCP (Model Context Protocol) servers. It scans your codebase for , maps findings to OWASP Agentic AI Top 10 and OWASP LLM Top 10 2025, generates a Trust Card with a letter grade, and provides a shareable badge for your README.

Local analysis, no OpenTrustBench telemetry. Dependency auditing uses the npm registry. No account required.

Quick Start

Install

| Channel | Command | Status |
|---------|---------|--------|
| npm | npm install -g /cli | v0.1.3 live (package) |
| npx | npx /cli scan . | No install needed |
| PyPI | pip install opentrustbench | v0.1.3 live (package) — requires Node 18+ and the npm CLI engine |
| Homebrew | brew tap eulogik/opentrustbench && brew install opentrustbench | v0.1.3 live (tap) |
| Docker | docker run --rm -v $(pwd):/workspace eulogik/opentrustbench scan . | v0.1.3 + latest live (Hub) |
| VS Code | code --install-extension eulogik.opentrustbench | v0.1.3 live (Marketplace) |
| GitHub Action | - uses: eulogik/opentrustbench-action@v0.1.3 (repo) | v0.1.3 live (Marketplace) |
| Source | git clone && cd OpenTrustBench && npm install | Always current |

What You Get

| Output | Description |
|--------|-------------|
| Trust Card | Machine-readable credential (opentrustbench/trust-card/v2) with grade, score, findings, permissions, and co |
| SARIF Report | Industry-standard format for GitHub Security tab integration |
| Markdown Report | Human-readable audit report for compliance and review |
| Grade Badge | Shareable S linking to a public report page (example, registry) |

Grade Badge

Embed your trust score in your README:

This tells buyers, auditors, and AI hosts that your agent has been .

How It Works

1. Scan — Point opentrustbench at a local directory, GitHub repo, or npm package
2. Grade — Receive a Trust Card with a letter grade (A–F) and security score (0–100)
3. Share — Embed your badge and link to the detailed report

Features

8 detection rules across shell injection, secret leaks, prompt injection, tool exploitation, and more

OWASP mapping — Every finding tagged to OWASP Agentic Top 10 (ASI01–ASI10) and LLM Top 10 2025 (LLM01–LLM10)

Permission manifest — Automatic extraction and scoping of agent permissions

Dependency audit — npm audit integration for lockfile-based detection

SARIF output — Nati Code Scanning integration

CI integration — --fail-on se gate for pipeline enforcement

GitHub Action — Drop-in composite action for workflows

GitHub URL detection — Scan any public repo by URL

Compliance helpers — Permission manifests and audit trails to attach to an EU AI Act review (e input, not a certification)

What the scanner is (and isn't)

8-rule static suite, OWASP-mapped (Agentic ASI01–ASI10, LLM LLM01–LLM10). Regex-based today — no AST yet.

attack is static-heuristic: it re-analyzes scan findings + permissions. No payloads execute.

eval is simulation mode: suites are parsed and ; nothing runs, costs/durations stay 0.

Not a certification. The Trust Card is a CI-grade credential, not a pentest or legal .

Expert reviews

Need a human pass over your results? Request a review — manual result triage plus a remediation plan, scoped per target. The CLI stays the product and is free forever.

| Tier | Price | What You Get |
|------|-------|-------------|
| CLI | Free | Unlimited local scans, Trust Card, badge, OWASP findings, SARIF, CI gate |
| Expert review | Scoped per target | E in CLI + manual triage, remediation plan, debrief call |
| Enterprise | Custom | E above + custom rules, e helpers, priority support |

Monorepo Structure

Build & Test

Standards

OWASP Agentic AI Top 10 (ASI01–ASI10) — Published December 2025

OWASP LLM Top 10 2025 (LLM01–LLM10)

EU AI Act Article 50 transparency obligations (enforceable August 2, 2026)

SARIF 2.1.0 for GitHub Code Scanning integration

License

Apache 2.0 © 2026 Eulogik

---

<p align="center">
<a href=" ·
<a href=" ·
<a href=" ·
<a href=" ·
<a href=" ·
<a href=" ·
<a href="
</p>

项目信息

开发语言：HTML

Star 数：13

技术要点

该项目属于安全软件方向的开源资源，代码结构清晰，适合安全研究人员学习参考。项目围绕Trust & security scanner for AI agents and MCP servers — OWASP-mapped findings, Trust Cards (A–F grades), SARIF + CI gate. Free CLI, zero telemetry.展开，提供了可运行的实现，对理解相关安全机制、工具原理与攻防技术具有较高的研究价值。

从技术实现角度看，此类项目的核心在于对底层原理的深入理解与工程化落地。开发者在研究过程中可以结合官方文档与社区讨论，梳理项目的初始化流程、核心模块职责与数据流转，逐步掌握其中的技术细节与设计思路。

使用与学习建议

1. 下载源码压缩包，解压到本地环境。

2. 根据项目类型准备对应的运行环境与依赖。

3. 按照 README 中的步骤完成配置与部署。

4. 在隔离的测试环境中运行验证，避免影响生产系统。

免责声明

本资源来源于开源社区，仅供安全研究、技术学习与合法授权范围内的攻防演练使用，请勿用于任何非法或侵害他人权益的用途。使用本项目产生的一切后果由使用者自行承担。