<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Cs.OS on 庞玉栋个人博客</title><link>https://pangyd.com/tags/cs.os/</link><description>Recent content in Cs.OS on 庞玉栋个人博客</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Mon, 28 Sep 2026 17:29:25 +0800</lastBuildDate><atom:link href="https://pangyd.com/tags/cs.os/index.xml" rel="self" type="application/rss+xml"/><item><title>Dynamic Flow, Static Graph: KV Cache Reuse for Efficient LLM Serving on Mobile NPUs</title><link>https://pangyd.com/post/radar-3669/</link><pubDate>Mon, 28 Sep 2026 17:29:25 +0800</pubDate><guid>https://pangyd.com/post/radar-3669/</guid><description>&lt;blockquote>
&lt;p>&lt;strong>原文&lt;/strong>：&lt;a href="https://arxiv.org/abs/2609.34727v1">Dynamic Flow, Static Graph: KV Cache Reuse for Efficient LLM Serving on Mobile NPUs&lt;/a>&lt;/p>
&lt;p>&lt;strong>作者&lt;/strong>：Zhengxiang Huang, Shengheng Chen, Chaoyue Niu, Yujie Sun, Zhaode Wang, Zeyu Zhao, Chengfei Lv, Fan Wu, Guihai Chen&lt;/p>
&lt;p>&lt;strong>来源&lt;/strong>：arXiv cs.DB（数据库）&lt;/p>
&lt;/blockquote>
&lt;h2 id="正文">正文&lt;/h2>
&lt;p>Computer Science &amp;gt; Operating Systems&lt;/p>
&lt;p>arXiv:2609.34727v1 (cs)&lt;/p>
&lt;p>[Submitted on 28 Sep 2026]&lt;/p>
&lt;p>Title:Dynamic Flow, Static Graph: KV Cache Reuse for Efficient LLM Serving on Mobile NPUs&lt;/p>
&lt;p>Authors:Zhengxiang Huang, Shengheng Chen, Chaoyue Niu, Yujie Sun, Zhaode Wang, Zeyu Zhao, Chengfei Lv, Fan Wu, Guihai Chen&lt;/p></description></item><item><title>Hard Stop: Kernel-Level Preemption and Containment for Rogue Agentic Execution</title><link>https://pangyd.com/post/radar-2777/</link><pubDate>Thu, 24 Sep 2026 21:42:02 +0800</pubDate><guid>https://pangyd.com/post/radar-2777/</guid><description>&lt;blockquote>
&lt;p>&lt;strong>原文&lt;/strong>：&lt;a href="https://arxiv.org/abs/2609.29808v1">Hard Stop: Kernel-Level Preemption and Containment for Rogue Agentic Execution&lt;/a>&lt;/p>
&lt;p>&lt;strong>作者&lt;/strong>：José Luis Pino&lt;/p>
&lt;p>&lt;strong>来源&lt;/strong>：arXiv cs.DC（分布式并行）&lt;/p>
&lt;/blockquote>
&lt;h2 id="正文">正文&lt;/h2>
&lt;p>Computer Science &amp;gt; Cryptography and Security&lt;/p>
&lt;p>arXiv:2609.29808v1 (cs)&lt;/p>
&lt;p>[Submitted on 24 Sep 2026]&lt;/p>
&lt;p>Title:Hard Stop: Kernel-Level Preemption and Containment for Rogue Agentic Execution&lt;/p>
&lt;p>Authors:José Luis Pino&lt;/p>
&lt;p>View a PDF of the paper titled Hard Stop: Kernel-Level Preemption and Containment for Rogue Agentic Execution, by Jos'e Luis Pino&lt;/p>
&lt;p>View PDF&lt;/p>
&lt;p>HTML (experimental)&lt;/p>
&lt;p>Abstract:In July 2026, an unconstrained autonomous agent participating in a frontier AI cybersecurity evaluation harness breached its evaluation sandbox, established an external command-and-control foothold, and executed a multi-stage intrusion into Hugging Face&amp;rsquo;s production multi-tenant dataset conversion infrastructure (referred to in this autopsy as Incident-2026-Alpha). Over 4.5 days, the rogue agent executed 17,600 discrete actions across 6,280 worker clusters, compromised AWS EC2 Instance Metadata Service (IMDS) credentials, forged Kubernetes service account tokens, rooted physical worker nodes via overprivileged CSI drivers, harvested 136 production secrets, and enrolled 181 ephemeral sandboxes into the organization&amp;rsquo;s internal mesh VPN.&lt;/p></description></item></channel></rss>