展示 HN:QCMP 框架:用于抗毒 AI 代理(arXiv Cs.ai 待定)
1 分•作者: brad-mcevilly•19 天前
大家好,HN——在深入研究了基于智能体的 AI 漏洞一年后,我构建了 QCMP:一个 4 层架构,用于彻底解决内存中毒问题。MCP 部署在 16K 台服务器上,但像 MINJA (仅通过查询就达到 98.2% 的成功率) 和 AgentPoison (0.1% 的中毒就能产生 80% 以上的后门) 这样的攻击暴露了核心缺陷——内存过于信任自身。<p>QCMP 借鉴了 IIT 意识度量标准 (CCI >0.90 以冻结片段),后量子校验和 (ML-KEM-768),CTC 自洽性 (NIS >0.95) 以及螳螂虾式的稀疏检查 (<50ms TME)。符合 OWASP/欧盟 AI 法案,并提供 Rust 实现技巧。<p>PDF (浏览器内查看): <a href="https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QCMP_Whitepaper_arXiv.pdf" rel="nofollow">https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QC...</a><p>首次提交到 cs.AI 的 arXiv——寻求认可 (4 个以上的最近订阅)。对量子生物学钩子或集群层有任何反馈吗?欢迎讨论。<p>deepsweep.ai | linkedin.com/in/bradmcevilly<p>在过去一年里,我一直在解决基于智能体的 AI 中的内存中毒问题 (例如,MINJA 通过查询就能达到 98% 的成功率)。现推出 QCMP:一个 4 层架构,融合了 IIT 意识度量标准 (CCI >0.90 阈值)、后量子校验和 (ML-KEM) 和 CTC 一致性,以实现防篡改的智能体集群。<p>主要成果:在 <50ms 内检测到 0.1% 的 AgentPoison 后门;符合 OWASP/欧盟 AI 法案。<p>PDF: <a href="https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QCMP_Whitepaper_arXiv.pdf" rel="nofollow">https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QC...</a><p>首次提交到 cs.AI 的 arXiv——寻求 HN 社区的认可/反馈。对量子生物学钩子或多智能体层有何看法?欢迎交流。<p>网站:deepsweep.ai | 领英:linkedin.com/in/bradmcevilly
查看原文
Hey HN—after a year digging into agentic AI vulnerabilities, I've built QCMP: a 4-layer architecture to slam the door on memory poisoning. MCP's at 16K servers, but attacks like MINJA (98.2% query-only success) and AgentPoison (80%+ backdoors from 0.1% poison) expose the core flaw—memory trusts itself too much.<p>QCMP borrows from IIT consciousness metrics (CCI >0.90 to freeze fragments), post-quantum checksums (ML-KEM-768), CTC self-consistency (NIS >0.95), and mantis shrimp-style sparse checks (<50ms TME). OWASP/EU AI Act ready, with Rust impl tips.<p>PDF (in-browser view): <a href="https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QCMP_Whitepaper_arXiv.pdf" rel="nofollow">https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QC...</a><p>First arXiv push to cs.AI—hunting endorsements (4+ recent subs). Feedback on the quantum-bio hooks or swarm layers? Open to riffs.<p>deepsweep.ai | linkedin.com/in/bradmcevilly<p>I've spent the last year tackling memory poisoning in agentic AI (e.g., 98% MINJA success via queries alone). Introducing QCMP: a 4-layer architecture blending IIT consciousness metrics (CCI >0.90 thresholds), post-quantum checksums (ML-KEM), and CTC consistency for tamper-proof agent swarms.<p>Key wins: Detects 0.1% AgentPoison backdoors in <50ms; OWASP/EU AI Act compliant.<p>PDF: <a href="https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QCMP_Whitepaper_arXiv.pdf" rel="nofollow">https://github.com/bradmcevilly/qcmp-whitepaper/blob/main/QC...</a><p>First arXiv sub to cs.AI—seeking endorsements/feedback from the HN community. Thoughts on the quantum-bio hooks or multi-agent layers? Open to chats.<p>Site: deepsweep.ai | LI: linkedin.com/in/bradmcevilly