Paired targets. Two-way validation. Every check must catch the vulnerable side and stay clean on the hardened side — anything untestable is marked untested, never "passed."
An open benchmark range and zero-dependency audit toolkit for the deployment and permission layer of AI agents and MCP servers — where tools execute, credentials live, and context gets poisoned. I audit what the model layer can't see: leaks, over-granted tools, unsafe defaults, session handling, CORS, error disclosure.
Not a scanner marketing page, and not model-side red teaming — that layer already has excellent free tools (promptfoo, garak, PyRIT, AgentDojo).
This practice is new, but the work behind it is public and reproducible. I don't invent testimonials; I list what already exists:
Everything below is automated, passive protocol-layer observation plus optional written reports — not a security certification and not a substitute for a qualified professional. Run it yourself, or ask for an assisted pass.
For solo developers and small teams who want to run the checks themselves: the same zero-dependency scanner, an OWASP mapping sheet covering both the MCP Security Cheat Sheet (12 risks) and the Agentic AI Top 10 (ASI01–ASI10), plus a self-service reporting template — anything it can't test is marked Untested, never "Pass". No appointment, no waiting. Buy now on Creem → (or email 535710934@qq.com if the link won't open).
For AI-agent / MCP project maintainers — and for enterprises who want a first look at their shadow-agent surface. Passive protocol-layer snapshot mapped to OWASP Agentic AI Top 10 (ASI01–ASI10), delivered as a sanitized HTML report. Sample: see a real one.
What the snapshot can't tell you: what each finding means for your business, what to fix first (down to files), and a retest after you ship fixes. Includes a one-page executive summary you can take to a meeting.
Your team ships continuously; I watch the regression surface after every release and tell you within 24h what broke and what got fixed.
The first three tiers answer "does this system have problems." This one answers a different question: how many agents are touching your data, what can each reach, and which to govern first. Deliverables: a full inventory (who installed what, when), a permission × data-source matrix, a prioritized remediation list, a one-page executive summary, and a compliance control-gap map. Read-only by design — no agent is executed, no business data is read, no configuration is changed.
Email 535710934@qq.com with the subject [Audit Request] <project name> and include: ① repo / deployment URL, ② confirmation that you own or maintain it, ③ the one thing that worries you most.
Free snapshot audits are limited to projects you own or maintain, one per project.
Enterprises: for a shadow-agent inventory (¥15,000–30,000, read-only), use the subject [Inventory Request] <company> and briefly describe where your agents run (endpoints / cloud / SaaS) and who should be the point of contact.
我审计 AI agent 的部署层与权限面——告诉你它会在哪里泄漏、哪里越权、哪里重复下单。不做模型侧红队，那一层已有免费工具做得很好。
免费快检（¥0，48 小时）：面向 AI agent / MCP 项目维护者的被动协议层观察，OWASP Agentic AI Top 10 (ASI01–ASI10) 映射，脱敏 HTML 报告。申请方式：邮件标题 [Audit Request] + 项目名，附仓库地址并确认项目归你维护。
付费档（¥3,000–8,000）回答体检报告回答不了的三个问题：这些问题对你意味着什么、先修哪个、修没修好。月度复扫（¥800–2,000/月）回答：改了这么多，还安全吗。
企业影子 agent 盘点（¥15,000–30,000，3–5 个工作日）：前三档回答的是"某个系统有没有问题"，这一档回答的是"我们公司到底有多少个 AI 在碰数据、每个能碰多少、先管哪个"。交付：全量清单（谁装的、什么时候装的）、权限 × 数据源矩阵、优先级整改清单、一页决策者摘要、合规控制缺口对照表。全程只读——不执行任何 agent、不读业务数据、不改任何配置。
靶场与 18 个审计工具已开源；审计方法论本身不开源。我证明不了的东西（反查命中率、未覆盖面）在仓库 README 里逐条写明——诚实是这套流程的第一交付物。
Most security pages sell certainty. I sell a process that documents what it can't prove — because the expensive failures in agent security come from checks that silently pass on things that were never tested.
If you deploy agents and want to know what your current posture actually is, email below — or buy the self-check toolkit on Creem and run it yourself today.
535710934@qq.com · [Audit Request]Everything on this page is reproducible from the public repository. Public research to date: FreeBuf bylined article (Oct 2026) · closed-loop third-party audit on a major MCP ecosystem · open-source toolkit aligned to OWASP Agentic AI Top 10 (ASI01–ASI10).