<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Sandbox on 兜忆轩 | Cedric's Compiler Notes</title><link>https://douyixuan.github.io/tags/sandbox/</link><description>Recent content in Sandbox on 兜忆轩 | Cedric's Compiler Notes</description><image><title>兜忆轩 | Cedric's Compiler Notes</title><url>https://douyixuan.github.io/images/papermod-cover.png</url><link>https://douyixuan.github.io/images/papermod-cover.png</link></image><generator>Hugo -- 0.146.0</generator><language>en-us</language><lastBuildDate>Wed, 16 Sep 2026 08:00:00 +0800</lastBuildDate><atom:link href="https://douyixuan.github.io/tags/sandbox/index.xml" rel="self" type="application/rss+xml"/><item><title>AI Native Digest 2026-09-16</title><link>https://douyixuan.github.io/posts/ai-native/2026-09-16/</link><pubDate>Wed, 16 Sep 2026 08:00:00 +0800</pubDate><guid>https://douyixuan.github.io/posts/ai-native/2026-09-16/</guid><description>今天关注两个 AI-native 系统方向：Google 把 Agent 安全从 prompt/regex 推进到运行时治理层，在模型外统一执行 prompt screening、语义策略和跨会话异常检测；T1 则展示了长时程 terminal agent 如何用真实 sandbox、可执行 verifier、dense process reward 与 MoE 路由一致性机制进行强化学习，把 coding agent 的能力提升从 inference-time harness 延伸到训练阶段。</description></item></channel></rss>