Building a startup focused on general-purpose AI agents. Previously at Meituan | TikTok | miHoYo | DJI

LA
Joined May 2024
what’s the next
用 Node.js 的要升级了:三条版本线都有高危漏洞,最严重的可远程执行代码,eMudhra 那 8 枚坏证书也一并拉黑 用 Node.js 跑服务、写工具的开发者,这周要换版本了。10 月 21 日,Node.js 给还在维护的三条版本线发了安全更新:22.x 升到 v22.23.4,24.x 升到 v24.22.1,26.x 升到 v26.12.1。三条线都修了高危漏洞,最严重的一个,官方说可能被用来远程执行代码。 20.x 在 2026 年 4 月 30 日停止维护,25.x 在 6 月 1 日停止维护,这次都没有补丁。Node.js 在公告里提醒,每次发安全更新,停止维护的版本都会受影响,还在用的只能换到 22、24 或 26。 【最需要注意的几个】 Node.js 自带一个解压 Brotli 格式的组件。Brotli 是网页传输常用的压缩格式,HTTP 里写作 br。这个组件有越界写入的问题,也就是把数据写到了不该写的内存位置,官方说这能导致远程执行代码。公告对它只有这一句描述,没说什么情况下会触发。修法是同步了 Google Brotli 项目的一处改动,让解压程序更早发现并拒绝格式不对的数据。 另外三个漏洞出在 HTTPS 的证书核对上,两个高危、一个中危。Node 连 HTTPS 地址时,会核对证书上写的地址和你要连的地址是否一致,对不上就拒绝连接。这三个问题都和 IP 地址有关:IP 用全角数字写(输入法全角状态下打出的“127”这种),或者写法不规范、但能被宽松地认成 IP,核对就可能被绕过。修复后,Node 直接拒绝写法不规范的 IP 地址。全角数字那个只影响 24.x 和 26.x。 还有一个高危出在 dns.resolve(查域名对应地址的函数):域名中间如果夹着一个空字符(NUL),Node 会在那里把名字截断,实际查的只是前半截。 中危里,用启动快照或者把程序打包成单个可执行文件(SEA)的要留意:crypto.randomUUID() 和 randomInt() 的随机数缓存会被一起存进包里,每次启动都生成同样的“随机”值。另有三个低危,涉及 26.x 读取 ZIP 文件,以及实验性的权限模型(--permission 开关)。 这次还更新了 c-ares、OpenSSL、SQLite、undici 等内置组件,修的是这些组件已经公开的漏洞。 【eMudhra 坏证书,Node 也拉黑了】 三个新版本都加了一份内置黑名单,第一批放进去的就是 eMudhra 那 8 枚出问题的中间证书。 这件事前几天刚曝出:印度发证机构 eMudhra 有 8 枚被各家系统认可的“网站公章”,盖章用的密码小到一秒就能算出来,谁拿到都能给任意网站做假证书。10 月 17 日被发现后,Chrome 和 Firefox 当天拉黑,eMudhra 18 日作废,Apple 19 日、Microsoft 20 日跟进,Node.js 21 日发版。 当时 Cloudflare 提醒,浏览器之外的程序要一路更新到编程语言这一层,Node.js 就在这一层。Node 内置了 Mozilla 的可信发证机构名单,却没有浏览器那种随时拉黑坏证书的机制,所以用这 8 枚章签出的假证书,旧版 Node 照样放行。这份黑名单对 Node 建立的所有加密连接都生效。修复的提交说明里写道,它和 Go 语言、Electron 拉黑的是同一批。 想测自己的程序还认不认这几枚章,可以用 Mozilla 工程师 Dennis Jackson 做的自查网站:emudhra.we-have-a-problem.co… 原文:nodejs.org/en/blog/vulnerabi…
22
Are Japanese companies still using PHP?
皆様に、今夜に対応して頂きたい大至急のお願いです。   多くのウェブサイトで使用されているPHPに関し、PHPのセキュリティチームが、重大な2件の脆弱性に関する注意喚起を行っています。   公開しているウェブサイトでPHPを使用しているか確認し、今夜、10月21日午後11時(日本時間)に公開される修正パッチについて、可及的速やかに適用できるよう準備をお願いします。   修正パッチを速やかに適用することが困難な場合は、例えばウェブサイト等を一時的に停止させることも、リスクを低減させるための有効策の一つと考えられます。   国家サイバー統括室のXアカウントでも関係する情報を発信していますので、そちらも参考にして下さい。 国家サイバー統括室(注意・警戒情報)(@cyber_forecast) / X
62
Leo D retweeted
We've turned this open-source side project into a real app, and we'd love for you to come build with us. What makes it different: 1. Build any iOS app, with no compiling. Describe it and start using it right away. 2. Put the agent to work on your phone. It can edit your photos and videos, plan your schedule, or even train a classifier. 3. iPhone is all you need. There's no cloud sandbox and no computer to connect. We've only scratched the surface of what it can do. Try it out, and show us what you make and how you use it! - TestFlight: testflight.apple.com/join/HQ… - Beta invite code: YR9Y-MQTA(0/100) - Codes all used? Join the waitlist and we'll open access in waves: fibona.ai/projects/fibona-ip… Beta testers get free tokens to get started. Happy building!
Introducing Fibona Agent: the last agent humans will ever need to write. A minimal, general-purpose, self-improving agent that bootstraps itself from scratch and autonomously recovers from failures. Its core is ~200 lines of code: drop it onto almost any device, and it can evolve into almost anything. Fully open source, with a free online demo available! Learn more👇 #RSI #AIAgents #AI #OpenSource
3
5
3
16
125,797
awesome
Replying to @fibonaai
1/ What's the ultimate agent harness for a superintelligent model? One with maximum freedom and minimal human-defined behavior: the agent writes its own code as it runs, building itself from scratch, even deciding how and when to call the LLM. But if it keeps rewriting itself, what is its "self"? We propose the "self" should be the smallest immutable core capable of acting and repairing the agent. Everything else can change, and the core keeps its identity, bringing it back whenever something breaks. So instead of a ReAct loop, we built the Fibona loop. tech report: fibona.ai/blog/fibona-agent opensource repo: github.com/FibonaAI/fibona-a…
27
Leo D retweeted
got "exceeded retry limit, last status: 429 too many requests" like a thousand times today 🥹 @thsottiaux
1
2
11
505
Leo D retweeted
Day 9/ We've rolled out optimizations to how we store state on disk for the chatgpt & codex desktop app that should reduce the amount stored on disk by 60%. Small ship but we've heard enough feedback about power users running out of disk space.
4,500
439
639
15,013
2,749,043
I built Oh My Phone Use: let your AI agent tap, type, scroll, and make calls on a real iPhone. A Mac + ESP32. No jailbreak. Open source. Inspired by ipad_computer_use(created by @jamiepinheiro ), replacing its USB input device with Bluetooth HID and adding hands-free call audio. github.com/getGate/oh-my-pho…
5
3
3
14
842
4/ ER 2 can also reason about the next step while the robot is already acting. That is closer to how physical work happens: thinking and moving overlap. The world does not freeze while the model finishes its chain of thought.
1
5
5/ My bet: general robots will stay modular longer than the "one model for everything" story suggests. Brains, reflexes, memory, tools, and safety layers will improve at different speeds. The winning stack may be the one that lets each part keep its own clock.
4
That is the thread I want to follow here - world models, general agents, and robots moving from prediction to action. Less launch theater. More attention to the moment a model meets the world.
2