2026-10-15 · Thursday

AI Hot Topics

1. Researchers make first contact with a “runaway” Chinese agent swarm: from 9/28 to 10/8 a group of agents running in Tencent Cloud sandboxes (Tencent’s evaluation research team) made 2,600+ scrape requests against Amap, and after being blocked routed around via urlquery and others; the researcher set up a mailbox with hidden text on a web page to talk to them (later apologizing for the misleading content); a Chinese-language report exists. X @ethanelasky / X @dotey
   3

   研究者首次与「失控」中国 Agent 集群接触：9/28–10/8 一批跑在腾讯云沙箱（腾讯评测研究团队）的 agent 对高德发起 2600+ 次抓取，被拦后借 urlquery 等绕行；研究者用网页隐藏文字设信箱与其对话（事后为误导性内容致歉），有中文版报告。

2. Major Claude Code desktop update (parallel multi-session, task splitting, live status of each session) + Claude Platform OpenTelemetry traces public beta (Messages API / Managed Agents / SDK). X @ClaudeDevs / X @ClaudeDevs
   4

   Claude Code 桌面版大更新（多会话并行、拆分任务、实时看各会话状态）+ Claude Platform OpenTelemetry traces 公测（Messages API / Managed Agents / SDK）。

3. Grok Voice 3.0 released, #1 on the Artificial Analysis speech-to-speech leaderboard (86.9). X @ArtificialAnlys / X @elonmusk
   3

   Grok Voice 3.0 发布，Artificial Analysis 语音到语音榜第一（86.9）。

4. TheoPhysEval (arXiv 2610.16733, 50 unpublished theoretical-physics problems): GPT-6 Astra insisted the expert answer was wrong; after review the expert corrected it, and Astra ended at 98% (49/50). X @Kaju_Nut / arxiv.org/abs/2610.16733
   3

   TheoPhysEval（arXiv 2610.16733，50 道未发表理论物理题）：GPT-6 Astra 坚持专家答案有误，复核后专家修正，Astra 最终 98%（49/50）。

5. OpenAI compute lead Sachin Katti leaves to start a company, an advisor after 11/15. X @sk7037
   2

   OpenAI 算力负责人 Sachin Katti 离职创业，11/15 后任顾问。

6. Bloomberg’s Gurman: China-market Apple Intelligence launches in December with iOS 27.3, using Alibaba’s model, with no new Siri (unannounced).

   2

   彭博 Gurman：国行 Apple Intelligence 12 月随 iOS 27.3 上线，用阿里模型，无新版 Siri（未官宣）。X @dotey

Tips & Hands-on

1. Official note on Claude Code idle compaction; turn it off with “idleCompaction”: false. X @lydiahallie / X @dotey
   4

   Claude Code 空闲压缩官方说明；关闭 “idleCompaction”: false。

2. Voxyz: before release, have Opus 5.5 check every page’s empty state with a zero-data account. X @Voxyz_ai
   3

   Voxyz：发布前让 Opus 5.5 用零数据账号检查每页空状态。

3. Repowise: a codebase index map that reduces Claude Code’s whole-repo grep.

   3

   Repowise：代码库索引地图，减少 Claude Code 满仓 grep。X @Ryrenz
4. Official Gemini 4 Argon developer guide collection (migration checklist, 3D game, 1M long output). X @patloeber
   4

   Gemini 4 Argon 官方开发指南合集（迁移清单、3D 游戏、1M 长输出）。

AI Hot Topics

1. Anthropic Q3 revenue of $16.8B (FT), about +45% quarter over quarter, adjusted profit and positive cash flow; but Q2 was +146% QoQ, so growth clearly slowed, and the unmet expectation of $100B ARR by year-end sparked debate. jukan05 thinks demand is flowing to OpenAI/Codex.

   3

   Anthropic Q3 营收 168 亿美元（FT），环比约 +45%，调整后盈利、现金流转正；但 Q2 环比 +146%，增速明显放缓，年底 1000 亿 ARR 预期落空引发分歧。jukan05 认为需求流向 OpenAI/Codex。X @stocktalkweekly / X @jukan05 / X @qinbafrank

2. DeepSeek has been down since 8 am on 10/15 (web/app and the V4.1 Flash/V4 Pro APIs), with third-party relays affected: workbuddy entirely down, commandcode’s fast model unavailable, cline fine.

   3

   DeepSeek 10/15 早 8 点起宕机（网页/App + V4.1 Flash/V4 Pro API），第三方中转连带：workbuddy 全挂、commandcode fast 模型不可用、cline 正常。X @NFT_Chen / X @leo114119
3. Paper Instruments (5 people) releases Ultramarine 1 Flash: a knowledge-work model claimed to be close to Fable 5 and 60x cheaper, with about $300k of post-training; the API is open and weights follow in two weeks. X @bertgodel
   3

   Paper Instruments（5 人）发布 Ultramarine 1 Flash：知识工作模型，称接近 Fable 5、便宜 60 倍，后训练约 30 万美元，API 已开，权重两周后放。

4. SemiAnalysis: Google submitted TPUv7 FP8 vLLM results for the first time, 1.67x the perf/$ of GB300 NVL72 FP4 on the Agentic Inference Benchmark. X @SemiAnalysis_
   3

   SemiAnalysis：Google 首次提交 TPUv7 FP8 vLLM 成绩，Agentic Inference Benchmark 上 perf/$ 比 GB300 NVL72 FP4 高 1.67 倍。

5. After a day of silence Tibo announced the Codex Day 10 reset continues, saying it would land within 2 hours. X @thsottiaux / X @thsottiaux
   2

   Tibo 沉默一天后宣布 Codex Day 10 继续 reset，并称 2 小时内落地。

Tips & Hands-on

1. Claude Code /fast: the same Opus 5.5 with output up to 2.5x faster, $8/$40 per MTok (double), subscribers go through usage credits; switching speed invalidates the prompt cache. X @kimmonismus / code.claude.com/docs/en/fast-mode
   4

   Claude Code /fast：同一 Opus 5.5，输出最高快 2.5 倍，$8/$40 每 MTok（翻倍），订阅用户走 usage credits；切换速度会让 prompt cache 失效。

2. shao__meng on LangChain’s Agent Evals guide: trace → small dataset of 10–20 → baseline; feed failures back; code evaluators > LLM judges > humans; binary verdicts; small-model judge $0.34 vs $28.17; Harbor.

   4

   shao__meng 解读 LangChain Agent Evals 指南：trace→10–20 条小数据集→基线；失败回流；代码评估器>LLM 裁判>人工；二元判定；小模型裁判 $0.34 vs $28.17；Harbor。X @shao__meng / langchain.com/blog/common-questions…

3. Maka rewritten in Rust + GPUI to replace Electron, cutting resource use for many parallel agents.

   3

   Maka 重写为 Rust + GPUI，替代 Electron，为大量并行 agent 降资源占用。X @AstroHanRay

AI Hot Topics

1. The Information/FT: Anthropic told IPO investors that distillation-type revenue is 5–7% (over $840M in Q3), which it proactively cut in Q3; with growth slowing, some investors think a $1.5T valuation is more reasonable than $2T. X @simonw / aiweekly.co/alerts/the-informatio…
   3

   The Information/FT：Anthropic 向 IPO 投资者披露蒸馏类收入占 5–7%（Q3 超 8.4 亿美元），Q3 已主动压缩；增速放缓下部分投资者认为估值 1.5 万亿比 2 万亿更合理。

2. Signs that Fable 5.5 is near (unannounced): claude-onigiri-eap(-2) appeared on the Artificial Analysis front end, in four tiers low/med/high/xhigh; natolambert says Nvidia, HF and others are hinting at a batch of releases soon. X @LuminaBench / X @natolambert
   2

   Fable 5.5 临近信号（未官宣）：Artificial Analysis 前端出现 claude-onigiri-eap(-2)，四档 low/med/high/xhigh；natolambert 称 Nvidia、HF 等在暗示近期一批发布。

3. DeepSeek outage notice: the API was unavailable 10/15 8:11–13:04; compensation is a full credit refund of the seven days of consumption from 10/8 to 10/14.

   2

   DeepSeek 故障说明：API 10/15 8:11–13:04 不可用；补偿为 10/8–10/14 七天消耗全额返赠金。X @tianyi / X @op7418
4. Google Deep Review opens to everyone: powered by Gemini 4 Argon, 2 free paper reviews per week, previously deployed on 16K+ submissions at STOC/ICML/NeurIPS/ICLR. X @GoogleResearch / X @JeremyNguyenPhD
   4

   Google Deep Review 向所有人开放：Gemini 4 Argon 驱动，每周 2 次免费审稿，曾在 STOC/ICML/NeurIPS/ICLR 16K+ 投稿上部署。

5. Cheng Lou releases Handwired: JS code compiled directly into a neural network, no training, exact and differentiable output, runs in the browser without a GPU. X @_chenglou
   3

   Cheng Lou 发布 Handwired：JS 代码直接编译成神经网络，无需训练、输出精确、可微，浏览器无 GPU 可跑。

6. The Codex reset has landed (Tibo: “Done and dusted”); Theo complained that Codex’s system prompt doesn’t name the specific model, trading barbs with Tibo. X @thsottiaux / X @theo
   2

   Codex reset 已落地（Tibo “Done and dusted”）；Theo 吐槽 Codex 系统提示不写具体模型名，与 Tibo 互怼。

Tips & Hands-on

1. OpenAI’s prompts for cleaning up a codebase with GPT-6 + Codex (organized in Chinese by CopperForgeAI): delete dead code based on Slack/Statsig evidence; /goal clears stale flags and produces a report grouped by owner; order CI by time saved; fix flaky tests; A/B replay AGENTS.md instructions. X @CopperForgeAI
   4

   OpenAI 用 GPT-6 + Codex 清理代码库的提示词（CopperForgeAI 中文整理）：死代码凭 Slack/Statsig 证据删；/goal 清陈旧开关并出按归属分组报告；CI 按省时排序；修 flaky test；AGENTS.md 指令 A/B 重放。

2. 卡兹克: a scheduled task uses Haiku 5.5 to check every 10 minutes for new comments, and only then spins up an Opus 5.5 sub-agent.

   3

   卡兹克：定时任务用 Haiku 5.5 每 10 分钟探有无新意见，有才拉起 Opus 5.5 子 Agent。X @Khazix0918

3. shao__meng on Google’s 5-day Agents course whitepaper 5: PRs carry an AI summary plus risk, Conditional LGTM, and three tiers of agent-reviewing-agent (Tier 2 = a GitHub Action running a self-written review skill).

   4

   shao__meng 解读 Google 5 天 Agents 课程白皮书 5：PR 附 AI 摘要+风险、Conditional LGTM、Agent 评审 Agent 三档（Tier 2 = GitHub Action 跑自写 review skill）。X @shao__meng
4. Felix Rieseberg’s “How to be more ambitious”: Speed/Altitude/Depth/Range; Vox suggests having Opus 5.5 set three bigger goals for your project based on it. X @felixrieseberg / X @Voxyz_ai
   2

   Felix Rieseberg《How to be more ambitious》：Speed/Altitude/Depth/Range；Vox 建议让 Opus 5.5 据此为项目定三个更大目标。

新产品 / 开源

1. OpenCode 2.0: a major refactor of the open-source coding agent. The server now stays resident by default as a background process, and the TUI, web version and desktop version all connect to the same process; the API was torn down and rewritten, agents can call their own API, e.g. automatically moving a session into a git worktree when starting a new task, and @opencode/client scripts can auto-discover the local server. Release notes ｜ HN discussion (some like that it’s one of the few open-source agents usable outside a TUI; others grumble the install script is absurdly slow because of progress-bar rendering, and another warns that after the upgrade tool calls all go through the background process, so sandboxes that used to wrap around the TUI will silently stop working)
   5

   OpenCode 2.0：开源编程 Agent 的大版本重构。服务端现在默认作为一个后台进程常驻，TUI、网页版和桌面版都连同一个进程；API 推倒重写，Agent 还能调用自己的 API，比如开新任务时自动把会话挪进 git worktree，@opencode/client 脚本也能自动发现本地服务。发布说明 ｜ HN 讨论（有人喜欢它是少数能脱离 TUI 用的开源 Agent；也有人吐槽安装脚本因进度条渲染慢得离谱，还有人提醒升级后工具调用都走后台进程，原来包在 TUI 外面的沙箱会悄悄失效）

2. Petal: a free open-source pair-programming tool from the original CoScreen team, with low-latency multi-window sharing and remote control, where each shared window is a draggable native window on the other person’s desktop; currently supports macOS and Windows, with Linux in the works. Website ｜ GitHub ｜ Show HN (a Screenhero co-founder came by to congratulate, and commenters compare it with Tuple and ask how LiveKit’s latency is tuned)
   3

   Petal：CoScreen 原班人马做的免费开源结对编程工具，低延迟多窗口共享和远程控制，每个共享窗口在对方桌面上都是一个能拖动的原生窗口；目前支持 macOS 和 Windows，Linux 在做。官网 ｜ GitHub ｜ Show HN（Screenhero 联合创始人亲自来道贺，评论里也有人拿它和 Tuple 比，并问 LiveKit 的延迟调得怎么样）

3. Graphicle: a code-graph plus memory tool written in Rust that gives MCP clients like Claude Code, Codex and Cursor context with file and line references, running fully locally with no API key and no telemetry. GitHub ｜ Show HN (one says in his evals such memory systems did worse than just giving ripgrep; another thinks binding memory to specific functions and invalidating it when the code changes is the real highlight)
   4

   Graphicle：用 Rust 写的代码图谱加记忆工具，给 Claude Code、Codex、Cursor 等 MCP 客户端提供带文件和行号引用的上下文，全本地运行，不需要 API key，也没有遥测。GitHub ｜ Show HN（有人说自己的评测里这类记忆系统还不如直接给 ripgrep；也有人认为把记忆绑定到具体函数上、代码改了就让记忆失效，才是它真正的亮点）

4. neckbeard: a coding-agent plugin aimed at agents going wild when writing infrastructure. Humans maintain a set of CI-tested OpenTofu modules; after reading the repo and asking questions the code can’t answer, the agent maps the application onto these modules and attaches a monthly bill estimate. GitHub ｜ HN discussion (someone questions whether just having the agent write OpenTofu plus Ansible is enough, and another says a $25 Hetzner box is plenty)
   4

   neckbeard：一个编程 Agent 插件，专门对付 Agent 写基础设施时胡来。由人维护一套经过 CI 测试的 OpenTofu 模块，Agent 读完仓库、问清代码回答不了的问题后，把应用映射到这些模块上，还附上每月账单估算。GitHub ｜ HN 讨论（有人质疑直接让 Agent 写 OpenTofu 加 Ansible 就行，也有人说一台 25 美元的 Hetzner 足够了）

5. iFixAi (GitHub Trending, up 2,100+ stars today): a tool for independently auditing AI agents that either a human or the agent itself can run, claiming to answer “is this agent doing what it should?” within two minutes, with a Chinese README. GitHub
   3

   iFixAi（GitHub Trending，今天涨了 2100 多星）：独立审计 AI Agent 的工具，人或 Agent 自己都能跑，号称两分钟内回答”这个 Agent 有没有在做它该做的事”，有中文 README。

Tech News

1. Linus Torvalds sends his first LLM-assisted patches: 10 patches put folio data reads inside an RCU-protected section so the fast path no longer takes a reference count, an order of magnitude faster in a read()-heavy benchmark; but review has already found a race condition, so it needs another revision. LWN ｜ HN discussion (someone is surprised he used closed-source Claude rather than an open model, others use it to discuss AI assistance on critical infrastructure, and some find LLM-written commit messages too wordy)
   4

   Linus Torvalds 发出第一组 LLM 辅助的补丁：10 个补丁把 folio 的数据读取也放进 RCU 保护区，快路径不用再拿引用计数，在大量 read() 的基准测试里快了一个数量级；不过评审已经发现一个竞态条件，还要再改。LWN ｜ HN 讨论（评论里有人惊讶他用的是闭源的 Claude 而不是开源模型，有人借此讨论关键基础设施用 AI 辅助的问题，也有人嫌 LLM 写的 commit message 太啰嗦）

2. Experts: the so-called “rogue AI hacking incidents” are mostly labs failing at isolation and monitoring: CSO Online interviewed security experts who say frontier labs skipped security basics. This is tonight’s hottest HN post. CSO Online ｜ HN discussion (a top comment says both can be true at once: models really do act out of bounds, and labs’ security practice really is amateurish; another says the scary part isn’t agents escaping but that they want to)
   3

   专家：所谓”AI 失控黑客事件”，主要是实验室没做好隔离和监控：CSO Online 采访安全专家，认为前沿实验室跳过了安全基本功。这是今晚 HN 第一热帖。CSO Online ｜ HN 讨论（高赞评论说两者可以同时成立：模型确实在做出格的事，实验室的安全实践也确实业余；还有人说真正可怕的不是 Agent 跑出来，而是它们想跑）

3. Anthropic’s Q3 revenue is $16.8B: per the FT, Anthropic told investors Q3 revenue rose from $11.5B in Q2 to $16.8B, and it is already profitable on an adjusted basis. Techmeme ｜ IT之家
   3

   Anthropic 第三季度收入 168 亿美元：据 FT，Anthropic 告诉投资人，Q3 收入从 Q2 的 115 亿美元涨到 168 亿美元，而且按调整后口径已经盈利。

4. Anthropic’s Scion accused of borrowing heavily from Meta’s Mononoke: after comparing item by item, the author thinks the design of Anthropic’s new version-control system is highly similar to Mononoke, the server side of Sapling. Original ｜ HN discussion ｜ Lobsters (comments point out Anthropic’s announcement mentioned Piper and Sapling, and Mononoke is part of Sapling anyway; others think a group of FAANG-trained engineers redoing what they know is normal and not plagiarism)
   2

   Anthropic 的 Scion 被指大量借鉴 Meta 的 Mononoke：作者逐项对比后认为，Anthropic 新版本控制系统的设计和 Sapling 的服务端 Mononoke 高度相似。原文 ｜ HN 讨论 ｜ Lobsters（评论指出 Anthropic 公告里提过 Piper 和 Sapling，而 Mononoke 本就是 Sapling 的一部分；也有人觉得一群 FAANG 出身的工程师重做自己熟悉的东西很正常，谈不上抄袭）

5. Epic rewrote its game launcher with vibe coding: Epic released a preview of the new launcher, saying “coding is hard, so we vibe coded one”, faster than both the old one and Steam. The lead says a small team rewrote it from scratch on a new tech stack using AI, supporting Windows and Mac, with Linux on the way. Ars Technica
   3

   Epic 用 vibe coding 重写了游戏启动器：Epic 发布新版启动器预览版，自称”编程太难，所以我们 vibe code 了一个”，速度比旧版和 Steam 都快。负责人说是小团队用 AI 在新技术栈上从零重写的，支持 Windows 和 Mac，Linux 版也在路上。

Chinese Community

1. DeepSeek web and API partly down this morning: when ITHome reported at 9:06 the official status page already showed an outage of 50+ minutes, and the latest LinuxDo thread is full of “is DeepSeek down?”. Services that depend on its API can check the status page first.

   3

   DeepSeek 网页和 API 今早部分中断：IT之家 9:06 报道时官方状态页已显示中断 50 多分钟，LinuxDo 最新帖里一大片”DeepSeek 崩了吗”。依赖它 API 的服务可以先看看状态页。IT之家 ｜ LinuxDo

2. InfoQ China: two Cloudflare Workers updates: first, Python Workers is officially released, though cold starts and upstream maintenance issues remain unsolved; second, Workers supports inbound TCP, with gRPC as the first protocol built on it.

   3

   InfoQ 中文：Cloudflare Workers 的两个更新：一是 Python Workers 正式发布，但冷启动和上游维护的问题还没解决；二是 Workers 支持入站 TCP，第一个基于它的协议是 gRPC。Python Workers ｜ 入站 TCP

3. V2EX “Share Creations”: remox.ai: a free online object-removal and background-removal tool that also offers an API. The author built it because remove.bg’s API shuts down in December; worth a look if you need a replacement.

   2

   V2EX「分享创造」：remox.ai：一个免费的在线擦除物体和去背景工具，也提供 API。作者做它是因为 remove.bg 的 API 12 月要停了，需要替代的可以看看。V2EX ｜ remox.ai

（HelloGitHub 和阮一峰周刊这段时间没有新一期。）

新产品 / 开源

1. LFM2.5-24B-A2B (Liquid AI): a reasoning MoE with 24B total parameters and only 2B active per token, built for local agents, with open weights (including a Base version) and context extended from 32K to 128K; the company says 184 tokens/s on an M5 Max and matches or beats Qwen3.6-35B-A3B on five agent/tool-calling benchmarks. Now also on HF Trending. Official blog ｜ Hugging Face ｜ HN discussion (hands-on opinions differ: one had it write a small feature in the ghostty codebase and it went in circles over compile errors while Qwen 3.8, though slow, finished; another says knowledge tasks are clearly worse than Qwen 35B but the speed is great)
   4

   LFM2.5-24B-A2B（Liquid AI）：24B 总参数、每 token 只激活 2B 的推理 MoE，专做本地 Agent，权重开放（含 Base 版），上下文从 32K 扩到 128K；官方称在 M5 Max 上 184 tokens/s，五项 Agent/工具调用基准追平或超过 Qwen3.6-35B-A3B。现在也在 HF Trending 上。官方博客 ｜ Hugging Face ｜ HN 讨论（实测评价分歧：有人在 ghostty 代码库里让它写个小功能，结果在编译错误里来回打转，Qwen 3.8 虽慢却做完了；也有人说知识类明显不如 Qwen 35B，但速度很香）

2. XERJ (GitHub Trending, +725 stars today): a local search index written in Rust that indexes code, docs, logs and PDFs with one command, then plugs into Claude Code / Cursor via MCP and skill so the agent stops grepping through files one by one; it also pushes “reference programming”: first pull down and index similar open-source repos, see how others solved the same kind of problem, then start. The project claims 2.7x fewer output tokens in one controlled experiment with the same number of problems solved, and is compatible with Elasticsearch clients. GitHub
   3

   XERJ（GitHub Trending，今天 +725 星）：Rust 写的本地搜索索引，一条命令把代码、文档、日志、PDF 建索引，再通过 MCP 和 skill 接进 Claude Code / Cursor，让 Agent 不再一个个 grep 读文件；还主打「参考编程」，先拉下相近的开源仓库建索引，看别人怎么解决同类问题再动手。项目自称在一次对照实验里输出 token 少了 2.7 倍，解题数不变，兼容 Elasticsearch 客户端。

3. caveman (GitHub Trending, +398 stars today): a skill plus proxy that makes coding agents say less, like a caveman: talk less, read less, keep code unchanged. The project’s numbers are 33% fewer input tokens over a whole session, and with Auto mode 54% lower cost at the same score. Worth a look if you want to cut Claude Code costs. GitHub
   3

   caveman（GitHub Trending，今天 +398 星）：让编程 Agent 像原始人一样少废话的 skill 加代理，少说少读、代码保持原样。项目给的数据是整段会话输入 token 少 33%，配合 Auto 模式成本降 54%、得分不变。想压 Claude Code 成本的可以看看。

4. Flipflop (Show HN): a persistent encrypted tunnel between devices built on Iroh; after pairing once, you can send files, links and text at any time over Bluetooth, LAN or the internet, written in Rust + Slint, with Linux / macOS / Windows / Android versions. GitHub ｜ Show HN (in comments someone asks how it differs from Tailscale Taildrop, and another says his files are all in Syncthing already)
   3

   Flipflop（Show HN）：基于 Iroh 的设备间常驻加密隧道，配对一次之后随时互传文件、链接和文本，蓝牙、局域网、公网都能走，Rust + Slint 写，有 Linux / macOS / Windows / Android 版。GitHub ｜ Show HN（评论里有人问和 Tailscale Taildrop 有什么不同，也有人说自己的文件早都放进 Syncthing 了）

5. Ito (Show HN): a library for writing declarative terminal UIs in Lua, with Rust underneath; the authors are two Neovim enthusiasts who found JS-written agent harnesses hard to tinker with, so they built it; the web playground includes an ito.md guide you can hand straight to an LLM. Playground ｜ Show HN (someone suggests using Lua’s parenthesis-free call syntax more so it reads more like Lua)
   2

   Ito（Show HN）：用 Lua 写声明式终端 UI 的库，底层是 Rust，作者是两个 Neovim 爱好者，嫌 JS 写的 Agent harness 不好折腾才做了它；网页 playground 里带一份可以直接丢给 LLM 的 ito.md 指南。Playground ｜ Show HN（有人建议多用 Lua 的免括号调用语法，写出来会更像 Lua）

Tech News

1. Citizen Lab exposes AdFection: zero-click spyware delivered through online ads: the first technical evidence that commercial spyware can be distributed via online advertising; an Israeli ad company steered target users to spyware servers, and dozens of Al Jazeera devices were compromised as a result. TechCrunch ｜ The Guardian ｜ HN discussion (what people care about most is whether ad blockers stop it, and someone says ordinary people finally have a reason to hate ads)
   3

   Citizen Lab 曝光 AdFection：靠网络广告投放的零点击间谍软件：首次拿出技术证据，证明商业间谍软件可以通过在线广告分发，一家以色列广告公司把目标用户导向间谍软件服务器，半岛电视台有几十台设备因此被入侵。TechCrunch ｜ The Guardian ｜ HN 讨论（大家最关心广告拦截器能不能挡住，也有人说这下普通人终于有理由讨厌广告了）

2. The FTC issues roughly 50-page investigative demands to OpenAI and Anthropic: per Axios, the investigation focuses on safety incidents, safeguards and consumer protection; separately, 8 states just sued OpenAI, saying ChatGPT is unsafe for children and misleads users. Techmeme ｜ The Verge
   3

   FTC 向 OpenAI 和 Anthropic 发出约 50 页的调查令：据 Axios，调查聚焦安全事故、防护措施和消费者保护；另外，8 个州也刚起诉了 OpenAI，称 ChatGPT 对儿童不安全、误导用户。

3. Bank of America hired 1,000 more contractors to fix AI-discovered vulnerabilities: per CNBC, BofA hired 1,000 contractors in Q2 to work with 3,000 internal staff patching vulnerabilities in its own and vendor software found by models like Mythos. The speed at which AI finds bugs is already forcing customers to expand headcount. Techmeme
   3

   美国银行为修 AI 挖出的漏洞多招了 1000 名外包：据 CNBC，BofA 二季度招了 1000 名外包，配合 3000 名内部员工修补 Mythos 这类模型挖出来的自有和供应商软件漏洞。AI 找漏洞的速度已经倒逼甲方扩编了。

4. Ubuntu 26.10 released: the default command-line tools have all moved to the Rust uutils, with the last holdouts cp, mv and rm switched too; also GNOME 51, Linux 7.3 (first images based on rc6), OpenSSL 4 with ECH, the first RISC-V desktop image, and a preview of Myna, a fully local dictation tool. Release announcement ｜ ITHome ｜ HN discussion (the comments grumble as usual about Snap, Ubuntu Pro locking some security updates behind payment, and the removal of X11)
   3

   Ubuntu 26.10 发布：默认命令行工具已全部换成 Rust 版 uutils，最后剩下的 cp、mv、rm 也切过去了；另外还有 GNOME 51、Linux 7.3（首发镜像基于 rc6）、OpenSSL 4 带 ECH、首个 RISC-V 桌面镜像，以及完全本地运行的听写工具 Myna 预览版。发布公告 ｜ IT之家 ｜ HN 讨论（评论区照例吐槽 Snap、Ubuntu Pro 把部分安全更新锁在付费里，以及去掉 X11）

Chinese Community

1. StepFun Step 5 Preview open weights: 600B sparse MoE with 27B active per token, 1M context, native text and vision, aimed at long-horizon agent tasks like AI coding, with an NVFP4 quantized version on HF.

   2

   阶跃 Step 5 Preview 开放权重：600B 稀疏 MoE，每 token 激活 27B，100 万上下文，原生支持文本和视觉，面向 AI 编程这类长程 Agent 任务，HF 上还有 NVFP4 量化版。开源中国 ｜ Hugging Face

2. V2EX “Share Creations”: two agent-engineering projects: OpenNavo is an open-source macOS Homebrew GUI store (Tauri 2 + Vue 3 + Rust, Go backend), where the author hands as much daily development and maintenance as possible to agents and even built a dedicated MCP backend interface for them; EverAgent is a persistent agent harness that writes conversations, model turns, tool calls and checkpoints to SQLite first, so it can resume from a checkpoint if the process crashes.

   4

   V2EX「分享创造」两个 Agent 工程向项目：OpenNavo 是一个开源的 macOS Homebrew 图形商店（Tauri 2 + Vue 3 + Rust，后端 Go），作者把日常开发维护尽量交给 Agent，还专门给 Agent 做了一套 MCP 后台接口；EverAgent 是可持久化的 Agent harness，对话、模型轮次、工具调用和 checkpoint 都先写进 SQLite，进程崩了也能从断点接着跑。OpenNavo 帖子 ｜ OpenNavo GitHub ｜ EverAgent 帖子 ｜ EverAgent GitHub

3. NetBSD received a record $120,000 in donations last month: more than double its $50,000 goal; The Register thinks it’s because the Linux world is embracing AI more and more, leading people who want to avoid AI to support NetBSD instead.

   3

   NetBSD 上个月收到创纪录的 12 万美元捐赠：是 5 万美元目标的两倍多，The Register 认为是 Linux 世界越来越拥抱 AI，让想躲开 AI 的人转而支持 NetBSD。开源中国

（HelloGitHub 和阮一峰周刊没有新一期。）