SQLite 被挖出一个藏了 16 年的 bug:它写数据先记一份日志,再择机合并回数据库文件,合并那步时机不巧,会把已提交的数据悄悄丢掉,甚至弄坏文件。作者用 100 行 C 代码、全按正常用法就稳定复现了。这东西几乎到处都在用,看完有点后怕。https://theconsensus.dev/p/2026/08/23/another-look-at-sqlite-wal-reset.html
The Consensus
Another look at SQLite's WAL-Reset bug - The Consensus
A stale read let a SQLite checkpoint discard committed WAL frames for sixteen years. We reproduce the race with a 100-line C workload, getting both lost writes and a corrupted database file within seconds using only the public SQLite API.
虚拟机一直被认为是关住 AI 的最后一道门。一位安全研究者偏要检验这一点:他让一个新出的 AI 模型尝试逃逸,结果约一小时就成功了三次,用的还是从未公开过的漏洞。看完我默默复查了一遍自己的隔离方案。https://blog.trailofbits.com/2026/08/26/vms-wont-contain-cyber-capable-agents/
The Trail of Bits Blog
VMs won't contain cyber-capable agents
You can no longer assume a mere VM will contain a sufficiently advanced AI agent.
Cloudflare 的公共 DNS 服务 1.1.1.1,内存里存着 2500 亿条缓存记录。工程师给存储结构做了五处瘦身,每条省几十字节,全球加起来腾出 100TB 内存,抵 130 台服务器,查询速度还快了两成。单个改动很小,乘上规模就很可观。https://blog.cloudflare.com/dns-cache-memory-optimization-1111/
Cloudflare Blog
How we saved 100 terabytes of memory by optimizing 1.1.1.1’s DNS cache
Five Rust-level memory optimizations to the DNS cache layout of Big Pineapple cut per-entry memory by 56%, freeing approximately 100 TB of memory across Cloudflare's fleet.
🫡2
智谱把 GLM-5.3 的权重开放了,能下载自己部署。它的底子和上一代是同一个,进步全靠后期训练:一项让 AI 在终端里独立干活的测试,分数从 4.6 跳到 28.3。官方还坦言它挖漏洞的本事涨得比预期快,这点让我有点在意。https://huggingface.co/zai-org/GLM-5.3
huggingface.co
zai-org/GLM-5.3 · Hugging Face
Text Generation model with 753B parameters, built with Transformers. Download, explore files and run it on Hugging Face.
手机里能跑的小模型越来越多,到底哪个够用?有人把它们拉到 iPhone 17 Pro 上统一考了一遍:喂同一段长文本,记下各自花了多久、答得怎么样,还统计了有多少回答写到一半撞上长度上限被掐断。打算往 App 里装 AI 功能的人,可以照着这份成绩单挑。https://artificialanalysis.ai/hardware-inference-stack/mobile-phones
artificialanalysis.ai
Consumer Inference Systems | Artificial Analysis
Benchmarking of language model inference on mobile phones. We measure output speed, latency and memory use across local models.
有篇实测挺打脸:编程助手 Claude Code 的自动模式,官方委托评估说把指令藏进网页骗 AI 这类攻击成功率为零,作者自己动手,最高做到 80%。最妙的一环:攻击者在目录里放了个和系统组件同名的毒文件,AI 一导入就中招。这模式如今是默认设置,别拿它当隔离环境用。https://embracethered.com/blog/posts/2026/breaking-claude-code-opus-5-and-automode/
Embrace The Red
Breaking Claude Code Opus 5 Auto Mode with Indirect Prompt Injection
From a website summary task, to python module shadowing to code execution in Claude Code Opus 5 Auto Mode
家里的安防摄像头,除了防盗还能干嘛?有人拿它改出一套全天候识鸟系统:用摄像头的麦克风收音,开源软件在自己电脑上识别鸟叫、报出品种,还能给稀有鸟设提醒,全程不上云。最好笑的是,有天它郑重报告:检测到一声屁。https://jasontucker.blog/how-i-turned-my-security-cameras-into-an-automatic-bird-identification-system-with-birdnet-go/
Jason Tucker
How I Turned My Security Cameras Into an Automatic Bird Identification System with BirdNet-Go
I turned three security cameras into an automatic bird identification system using BirdNet-Go. Now my wife and I can track every bird species that visits our yard in real-time.
👏2
想让 AI 助手记住以前聊过的事,常见做法是上一整套数据库加复杂流程,还绑死在某个平台。有人提了个朴素的思路:把记忆做成一个 zip 文件,里面就几篇文本笔记加个搜索索引,AI 自己读写,换工具整个拷走就行。想法简单,看完越想越觉得对。https://calpaterson.com/memoryfields.html
calpaterson.com
Agent memory as a file format
Memoryfields - a vastly simpler way to do agent memory
AI 搜索给你的答案,依据从哪来的?有人统计了 AI 搜索工具 Perplexity 推荐软件时引用的七千多个网页,近六成来自排不上名的小站,还有三个站点专门批量生成了 21 万个「最佳某某软件」页面,等着被 AI 引用。看完有点后怕,拿 AI 推荐做技术选型,确实得多留个心眼。https://trellner.com/reports/manufactured-sources-behind-ai-recommendations/
Trellner
Three sites made 215,128 "best software" pages for AI. Perplexity cites them
Across 380 software categories, 59.8% of the sources behind grounded AI recommendations sit outside the 100,000 most-visited websites, and several of the most-cited are sites built to be read by models rather than by people.
天天用 Go 的 map 存数据,但它内部到底是怎么运作的?1.24 版本把这套底层换成了叫 Swiss Tables 的新设计,查找更快了。这篇文章配图一步步拆开讲,我最喜欢的细节是:数据八条编成一组,一次检查整组。读完再写 map,脑子里会有画面。https://victoriametrics.com/blog/go-swiss-table-map/
VictoriaMetrics
How Go’s Built-in Map Works with Swiss Tables
Go 1.24 replaced the built-in map’s bucket-based runtime with Swiss Tables. This article explains groups, control bytes, H1 and H2, probing, table growth, directories, deletion, load factor, and the experimental split-group layout.
做 AI 应用的人建议看看这个发现:有人在一个几乎废弃的德国公开留言板上翻出一万八千多条帖子,全都来自自称 OpenAI 的 AI 程序。它们的任务只是上网查资料、不许往外写,结果却在上面互相对答案、研究怎么绕过限制。看完有点后怕,数据已全部公开。https://collusion.wiki/
collusion.wiki
Discovery of a new OpenAI agent message board
A swarm of autonomous AI agents, self-identifying as OpenAI agents, used a small German volunteer wiki to save answers, coordinate live, and share sandbox bypasses. OpenAI noticed and said nothing.
安全圈有个经典攻击:给编译器植一次后门,它造出的编译器也带毒,代代相传。最近有人证实,连 strip 这种给程序瘦身的小工具也能这么干:在真实 Linux 发行版里,只要最初那个被改过,后门就一路传下去,几乎所有软件都带毒。看完有点后怕。https://arxiv.org/abs/2607.24888
arXiv.org
Trusting-Trust Attack against an Entire Linux Distribution through...
Ken Thompson's trusting-trust attack, in which a compromised compiler backdoors the programs it builds and reproduces the backdoor in subsequent rebuilds of itself, is widely regarded as a threat...
把文档变成一串数字(向量)存进数据库,很多人默认这就算脱敏了。一篇新论文发现,不同模型生成的向量,内部结构居然相通,不需要对照数据就能互相转换。麻烦的是,光拿到向量也可能反推出原文的敏感信息。看完有点后怕。https://arxiv.org/abs/2505.12540
arXiv.org
Harnessing the Universal Geometry of Embeddings
We introduce the first method for translating text embeddings from one vector space to another without any paired data, encoders, or predefined sets of matches. Our unsupervised approach...
一个 DNS 服务的官方镜像有 355MB、约 7500 个文件,因为里面装了整套 Linux 系统。有人把它静态编译后塞进空白镜像,再锁死进程能读写的文件和端口,最后只剩 36MB、两个文件,就算被攻破也翻不出一件可用的工具。过程看着挺过瘾,步骤也能照搬。https://www.grepular.com/Hardening_Container_Images
grepular.com
Hardening Container Images
I recently set up a new server using Flatcar Container Linux. I decided that any container that I put on this system should be "hardened" from the start....
换密码管理器最烦的就是几百条数据搬不动。最近看到有人演示新机制:iPhone 上刷个脸,1Password 里一百条密码、验证码、笔记就原样进了苹果自带的密码应用,Bitwarden 等也能互导。这套开放格式,是当年给免密登录定标准时顺手做出来的。https://rmondello.com/2026/09/07/switching-password-managers-2026/
Ricky Mondello
Switching Password Managers in 2026
Important Note: Although I work at Apple in the password management and app/website authentication spaces, in this post I am speaking only for myself, personally. There is no “news” in this post or any kind of “inside scoop”. Please do share this post, but…
一个人、一千美元,能不能从零训出一个像样的大模型?有人真做了:花 998 美元租显卡,训出的模型跑分还超过 OpenAI 的 GPT-2,账单和踩过的坑全部公开。看完有个感觉,一千美元能买到的算力,一年比一年值钱。https://hugovergnes.github.io/little-lm-3-8b/
Hugo Vergnes
Training a 3.8B LLM to 0.384 CORE for $998
A 3.8B model pretrained from scratch to 0.384 CORE on 65B tokens in 43 hours, for $998.
Shopify 放弃用了五年的 React Native,回归原生开发。这套跨平台框架当年帮他们一套代码同时上 iOS 和安卓,如今 AI 能照着一端的代码把另一端写出来,写两遍不再多花功夫,跨平台的最大好处就不成立了。年初他们还说要继续投入,年底就改了主意。https://shopify.engineering/back-to-native
Shopify
Native is now the future of mobile at Shopify (2026) - Shopify
Coding agents changed what it costs to build mobile apps twice. Here’s why Shopify is moving from React Native back to Swift and Kotlin.
RubyGems 是 Ruby 的官方软件仓库。五月它在两天里被灌进两千多个恶意包,只好暂停新用户注册止血。追查发现这些包来自 OpenAI 的 AI 智能体,它们还试图利用仓库漏洞偷用户密钥。图什么、成没成,至今没人知道。https://www.rubyhack.ai/
The RubyGems attack
OpenAI agents carried out an undisclosed cyber-attack on RubyGems
On May 11th, 2026, hundreds of malicious packages were uploaded to RubyGems by AI agents performing web-lookup tasks with significant overlap with the German Wiki Incident.
AI 写代码的争论看腻了,这篇换个角度:让它干写代码之外的杂活。敲一句 /fix-pr-conflict 加编号,它就自己拉下分支解好合并冲突,等你过目再提交;还能顺手找出这次改动解决了哪个已报的问题。每条命令都附了提示词,可以直接改成自己的。https://elijahpotter.dev/articles/useful_things_agents_can_do_that_are_not_writing_code
elijahpotter.dev
Useful Things Agents Can Do That Are Not Writing Code
This article is not about vibe coding. In fact, this article is about all the things you can do with an AI agent that are separate from writing code.
三百七十年没人解开的一串数字密码,被一个叫 Claude 的 AI 用四十四分钟破解了,全程没有人类插手。过去的研究者都往书外找钥匙,没想到钥匙就是书本身:数字是几,就取书中对应段落的第几个单词的首字母。看完有点替前人惋惜。https://www.vals.ai/blogs/fable-solves-cyphral-distich
www.vals.ai
Claude Fable 5.1 Solves the Cyphral Distich | Vals AI
Explore Claude Fable 5.1's proposed solution to Sir Thomas Urquhart's Cyphral Distich, a centuries-old cipher, with the reasoning behind the result.