<?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>AI 资讯 on Joey's AI Lab</title><link>https://www.399041.xyz/tags/ai-%E8%B5%84%E8%AE%AF/</link><description>Recent content in AI 资讯 on Joey's AI Lab</description><generator>Hugo</generator><language>zh</language><copyright>Joey</copyright><lastBuildDate>Fri, 09 Oct 2026 02:17:00 +0800</lastBuildDate><atom:link href="https://www.399041.xyz/tags/ai-%E8%B5%84%E8%AE%AF/index.xml" rel="self" type="application/rss+xml"/><item><title>一条提示词劫持 AWS 账户内全部 AgentCore 智能体:AgentCorruption 复盘</title><link>https://www.399041.xyz/posts/zenity-agentcorruption-aws-agentcore/</link><pubDate>Fri, 09 Oct 2026 02:17:00 +0800</pubDate><guid>https://www.399041.xyz/posts/zenity-agentcorruption-aws-agentcore/</guid><description>Zenity 披露针对 Amazon Bedrock AgentCore 的攻击链 AgentCorruption:聊天权限即可经 IMDS 窃取凭证,默认角色跨全部智能体,长期记忆可被投毒。</description></item><item><title>AI 不是笨,是瞎:企业级 Agent 记忆的三个设计决策</title><link>https://www.399041.xyz/posts/tessl-agent-memory-design/</link><pubDate>Fri, 09 Oct 2026 02:14:00 +0800</pubDate><guid>https://www.399041.xyz/posts/tessl-agent-memory-design/</guid><description>Tessl 工程博客:存主张而非片段、时间和缺失要显式、权限在模型运行前决定。消融实验里,禁止说「不知道」把 30 次诚实拒绝变成了 30 个编造的答案。</description></item><item><title>Arena 完成 2 亿美元 B 轮,顺手发布 Alignment Index</title><link>https://www.399041.xyz/posts/arena-200m-series-b-alignment-index/</link><pubDate>Fri, 09 Oct 2026 02:11:00 +0800</pubDate><guid>https://www.399041.xyz/posts/arena-200m-series-b-alignment-index/</guid><description>Arena 宣布 2 亿美元 B 轮(估值 31 亿美元,Felicis 领投),年化收入破亿,同步推出衡量 Agent 是否安全、真实、不越界的 Alignment Index。</description></item><item><title>定时跑的 Claude:Managed Agents 每日简报智能体的六个组件</title><link>https://www.399041.xyz/posts/claude-managed-agents-daily-brief/</link><pubDate>Fri, 09 Oct 2026 02:08:00 +0800</pubDate><guid>https://www.399041.xyz/posts/claude-managed-agents-daily-brief/</guid><description>Anthropic 发了 Claude Managed Agents 的定时简报参考实现:来源、目标、智能体、计划、记忆、护栏六个组件,书签代替时间窗口,发送确认后才推进状态。</description></item><item><title>Goodfire 给 Kimi K3 和 GLM 5.3 装上生产级网络安全监控:探针 + Judge 级联</title><link>https://www.399041.xyz/posts/goodfire-production-cyber-monitors/</link><pubDate>Fri, 09 Oct 2026 02:05:00 +0800</pubDate><guid>https://www.399041.xyz/posts/goodfire-production-cyber-monitors/</guid><description>Goodfire 用激活探针做廉价初筛、LLM judge 只看可疑交互的级联监控:召回 93%、成本降约 50 倍、延迟从 2750ms 到 42ms,已部署进生产。</description></item><item><title>OpenAI 封禁一俄一伊两个隐蔽影响行动,首个 Category 5 出现</title><link>https://www.399041.xyz/posts/openai-disrupts-dark-clark-bogus-bylines/</link><pubDate>Fri, 09 Oct 2026 02:02:00 +0800</pubDate><guid>https://www.399041.xyz/posts/openai-disrupts-dark-clark-bogus-bylines/</guid><description>OpenAI 封禁俄罗斯 Dark Clark(首个 Category 5)和伊朗 Bogus Bylines(Category 4)两个假前台影响行动:AI 主要提升规模与语言流畅度,假前台被曝光后依然脆弱。</description></item><item><title>能付钱的 Agent:LangChain 用 Stripe Link 和托管深度智能体搭了个 Restock</title><link>https://www.399041.xyz/posts/langchain-restock-agent-payments/</link><pubDate>Fri, 09 Oct 2026 01:59:00 +0800</pubDate><guid>https://www.399041.xyz/posts/langchain-restock-agent-payments/</guid><description>LangChain 开源 Restock:Slack 里的办公用品采购 Agent,搜真实商品、建购物车、经 Stripe Link 钱包付款。核心设计是模型永远接触不到支付机密。</description></item><item><title>Nano Banana 2.1 评测:文生图编辑双榜第 4,1K 图价格砍半</title><link>https://www.399041.xyz/posts/nano-banana-2-1-eval/</link><pubDate>Fri, 09 Oct 2026 01:56:00 +0800</pubDate><guid>https://www.399041.xyz/posts/nano-banana-2-1-eval/</guid><description>Artificial Analysis 评测 Google Nano Banana 2.1:单一模型同时做文生图与图像编辑,双榜第 4,1K 分辨率 0.0336 美元一张,是前 7 名里最便宜的。</description></item><item><title>Anthropic 2026 使用政策更新:欺骗性活动独立成章,武器禁令写细</title><link>https://www.399041.xyz/posts/anthropic-2026-usage-policy-update/</link><pubDate>Fri, 09 Oct 2026 01:53:00 +0800</pubDate><guid>https://www.399041.xyz/posts/anthropic-2026-usage-policy-update/</guid><description>Anthropic 发布 2026 使用政策更新,11 月 12 日生效:欺骗性活动整合成新章节、选举定向全面禁令取消、武器与监控条款写细、新增禁止持续虐待模型。</description></item><item><title>Google 开源 AQuA:常驻生产环境、给 Agent 做体检的质量智能体</title><link>https://www.399041.xyz/posts/google-aqua-ambient-quality-agent/</link><pubDate>Fri, 09 Oct 2026 01:50:00 +0800</pubDate><guid>https://www.399041.xyz/posts/google-aqua-ambient-quality-agent/</guid><description>Google 开源 AQuA:五阶段流水线采样生产轨迹、聚类验证失败模式、定位到代码行,实测单智能体每会话成本约 0.007 美元。</description></item></channel></rss>