reddit · SideProject · 11-01 20:50 45 分 ← 返回列表

I built nomosguard: a fail-closed security gate where the model explains the decision and a hash-chained ledger makes it

The bug I did not have: a log I could not tamper with, until someone rewrote the whole thing I spent a while trusting "tamper-evident" as a property of the file. My ledger was append-only, every entry hashed the previous one, and the tests proved that editing a field, reordering lines, truncating the tail or forging a link all got rejected at load. Then I wrote the adversarial case properly and found the hole: if you rewrite every record AND recompute every entry hash AND recompute every li

五维评分

付费意愿

需求真实度

个人可落地

竞争空间

新鲜度

AI 分析

痛点：AI Agent 和 MCP 工具的审计日志其实无法真正防篡改——只要整份文件被重写并重新计算全部哈希，日志自身仍然“一致”，因此需要把锚点放在文件之外，并对每次放行或拦截留下可解释的记录

目标用户：构建 MCP、LLM Agent 的独立开发者；以及需要审计留痕与合规举证的安全团队（金融、医疗、企业内控）

证据："if you rewrite every record AND recompute every entry hash AND recompute every link AND rewrite the header, the file is internally consistent again. My chain verifies that records agree with each other. It has no way to notice that all of them changed together." / "A digest mismatch raises LedgerFileError instead of quietly handing you a different history."

变现：核心库保持 Apache-2.0 免费；商业层做托管式锚定与合规 SaaS：每日将 ledger_digest 写入外部不可变存储（S3 Object Lock / 透明日志）并按 agent 会话计费，另加团队审计报表导出、SSO 与私有部署授权

MVP：两周内做一个“托管锚定 + 审计可视化”薄层：CLI 命令把 ledger_digest 推送到 Object Lock 存储或公共透明日志，再配一个简单 Web 面板按时间线展示策略决策与拦截原因，并支持一键导出可验证的审计报告

竞品：Langfuse、LangSmith、Helicone、Arize Phoenix、immudb、Trillian

摘要：作者发布 nomosguard 0.7.10（Apache-2.0、Python 3.10+、无运行时依赖），指出可防篡改日志的致命漏洞：整体重写并重算哈希后文件仍自洽，因此把 digest 锚定在文件外，并通过 MCP 会话边界做 fail-closed（失败时默认拒绝）放行。帖子是纯发布，未出现任何付费意愿或收入数据，属于技术洞察型机会，商业化需依赖后续的合规审计需求验证。