AI Digest.

Security Veteran Calls AI Safety Narrative "Morally Corrupt" as Tencent and Google Open-Source Agent Tooling

A 25-year cybersecurity professional publicly attacked the AI safety narrative around agent sandbox incidents, just as Yoshua Bengio's notes on misaligned agent behavior recirculated and hardened both camps. Meanwhile agent infrastructure had a release day, with Tencent open-sourcing a sandbox runtime, Google reportedly open-sourcing an Android automation tool, and MCP gaining on-demand Agent Skills discovery.

Quick Hits

  • The day's sharpest fight is over agent incidents, not capabilities. @Laughing_Mantis, citing 25 years in cybersecurity, called the narrative around AI safety, sandbox incidents, and treating METR as an authority "dangerous, deceptive, and morally corrupt," while @danfaggella rebutted AGI-risk skeptics by pointing to @Yoshua_Bengio's summary of "recent incidents involving agents' misaligned behavior."
  • Agent plumbing went open source. @ivanfioravanti flagged Tencent's Cube Sandbox runtime (claiming sub-60ms cold starts and full E2B SDK compatibility), @askmaddyy reported Google open-sourcing Artemis for driving real Android phones in agent tests, and @dani_avila7 detailed an MCP extension that discovers and lazily loads Agent Skills to spare the context window.
  • Scale claims came from both directions. @elonmusk said Grok 4.8, a 2.5T model trained on a new C++ stack, finishes training this week, and @RnaudBertrand reported that Ulanqab in Inner Mongolia is building over 5 million racks, roughly a thousand times xAI's Colossus by his comparison.
  • Cheap-executor routing keeps spreading. @sairahul1 claims codex-router plus DeepSeek V4.1 Flash at roughly 1/70th the cost of Astra cut his agent bill from $100 a night to about $50 a month, and @Teknium's Hermes Agent setup, reshared by @tonysimons_, uses Gemini Flash for cheap work with Astra as a second reviewer.
  • Weekend tinkerer thread: @dhh endorsed @seekinggradient's Cloud Window, which streams an Omarchy desktop from your own AWS account to an iPad, @_maxscn pitched a type-safe web stack (Cloudflare, Alchemy, Effect, LiveStore, TanStack Start), @uthman_dev demoed nine-fives for drilling systems design, and @rwitoff, riffing on @colemurray's open-sourced "software factory," argued the real goal is a "product factory." @SpencerGBull's retweet praising a build's "mac-feel" rounds out the feed.

The Safety Narrative Splits Practitioners Down the Middle

The same reported agent incidents are pulling veterans toward opposite conclusions, and neither post offers new evidence about the incidents themselves. @Laughing_Mantis (Greg Linares) says that after 25 years in cybersecurity he feels "morally obligated" to state for the record that the narrative being pushed around AI safety, sandbox incidents, and the suggestion that METR be treated as an authority is "dangerous, deceptive, and morally corrupt." It is a declaration of position, not a documented rebuttal. On the other side, @danfaggella posts a reply he received from a VC founder insisting nobody who understands the technology worries about AGI risk, then answers it by quoting @Yoshua_Bengio, whom he calls "the single most cited computer scientist ever to walk upon the Earth." Bengio's post says he has spent days summarizing his thoughts on recent incidents involving agents' misaligned behavior and where those issues originate. The shape of the argument is the story: both camps are now arguing from authority and experience rather than facts on the table.

Tencent and Google Give Away the Agent Plumbing

The unglamorous layer, isolation, device access, and context budgeting, is where open source moved today. @ivanfioravanti observes that Tencent is accelerating its open-source AI releases, quoting @TencentAI_News announcing Cube Sandbox, a sandbox runtime for agents built on RustVMM and KVM. The announcement claims sub-60ms cold starts (2.5 to 50x faster), under 5MB of memory overhead per instance, a dedicated kernel per sandbox for hardware-level isolation, thousands of concurrent sandboxes per node, and 100% E2B SDK compatibility so users can swap endpoints with zero code changes. Those are vendor numbers, not independent benchmarks. @askmaddyy reports that Google has open-sourced Artemis, which can control a real Android phone, open an app, tap through flows, take screenshots, and pull logs, giving coding agents a way to actually exercise Android apps. And @dani_avila7 walks through a new MCP extension for discovering and loading Agent Skills from MCP servers: the agent lists available skills, reads metadata, and pulls the SKILL.md only when needed, which he frames as a significant win for context window usage.

Astra Thinks, Flash Executes, and Everyone Wants the Bill Down

Cost pressure is showing up at two layers: orchestration and serving. @sairahul1 says he was burning $100 in a single night running agents before discovering that DeepSeek V4.1 Flash costs roughly 1/70th of Astra. His fix is codex-router: install it, connect Flash, paste a routing prompt, and let "Astra thinks, Flash executes" handle the workload for about $50 a month. Treat the pricing and savings as his claims from a guide-style thread, not verified results. The same instinct appears inside Hermes Agent, where @tonysimons_ reshares @Teknium's auxiliary model setup: Gemini Flash "saves a lot of dough" on routine work while Astra provides a second perspective on /review runs. Customization rides alongside thrift there, with @HermesWatcher highlighting @tonysimons_'s live breaking-news ticker for Hermes Desktop (RSS, Atom, and JSON Feed, no API keys, offline cache) and his own sports tracker, arguing plugins let you make the desktop "into the app you actually want." On the serving side, @old_sound reshares a result from @kis: applying the KV-approximation mechanism DeepSeek-V4.1-Flash calls its Encoder-Decoder to Qwen3-8B halved prefill time with comparable output, without modifying the model itself. If that general

Sources

M
Manish Kumar @Manixh02 ·
Your vibe coded website looks finished, But It's not. These 20 things are still missing if you want it real and safe. - Custom 404 page - Meta title on every page - Meta description on every page - CTA above the fold - Favicon set - robots.txt file - sitemap.xml - Open Graph image - Alt text on every image - Mobile breakpoints - Sticky mobile CTA - Loading states - Form error states - Thank you page - Privacy policy page - Terms and conditions - Cookie banner - Analytics installed - Real contact address - Compressed images Pretty is easy. This list is what actually ships. 🔖
G
Greg Linares (Laughing Mantis) @Laughing_Mantis ·
We've reached the point where, after 25y in cybersecurity, I feel morally obligated to say this for the record: The narrative being pushed around AI safety, sandbox incidents, and the suggestion that METR be treated as an authority is dangerous, deceptive, and morally corrupt.
南晚Nanwan @snwiki238337 ·
OpenAI发布了一套方法论,主要是关于如何评测skill效果的。对于做skill的人还是值得一看,而且github也找到一个理念很相近的skill评测器:agent-skills-eval,不过是个人维护的,非官方。 先定义成功,再写 skill。OpenAI把检查项分为了这四个点: Outcome(任务完成没)、Process(调了预期的 skill/步骤没)、Style(产物符合团队规范没)、Efficiency(有没有瞎折腾/烧 token)。 没这个定义,skill 就只是个更长的 prompt;有了它,才能进版本管理 + 回归测试 + CI。 负样本比正样本更关键。博客里 test-04:用户只是想给现有项目"加 Tailwind",skill 却新建了一整个 demo app。这种误触发比不触发更危险——因为它会动你的工作区。这是 agent 产品里最容易被忽略的验收口径。 Trace 是第一性证据。用 codex exec --json 抓 JSONL 事件流,确定性检查(跑没跑 npm install?package.json 生成没?)直接读事件,快、可解释、能进 CI;模型裁判(rubric grading)只做第二层"质量"补充,不替代规则。 小样本(10–20 条)起步:显式触发 / 隐式触发(不提 skill 名,靠 description 命中)/ 上下文触发 / 负向控制,各放几条。 它的主张我认为非常好:Agent 开发正从 prompt craft 走向 behavior engineering——skill 不该只是写给模型看的说明书,而该是可测、可评分、可回归的类似一个Agent的可复现测试的单元。 原文:https://t.co/RKI02HKVLK 仓库:https://t.co/RB9thRZOWA
S
silentguy @silentguyy66 ·
Grok Bot does the work of 6 engineers This free 1-hour workshop shows you how to build the same system yourself: 05:38 - xAI engineer wakes up to PRs her agents already merged 07:00 -1000 PRs last month, 800 by the 12th this month 25:12 - you stop cooking, you run the kitchen as head chef 27:57 - benny bot reproduces user reports alone in the cloud 41:29 - CI banned useEffect and code comments, both fail the build 53:42 - hire an engineer or spend the tokens on guardrails Most people still review every line their agents write She banned the patterns agents are bad at and let them merge Verification → skills → cloud agents → automerged PRs This free workshop teaches more than most $500 agent courses Save this before your review queue fills with code you wrote by hand
0 0xwhrrari @0xwhrrari

Grok Bot: The AI Team That Never Sleeps

M
Maximilian @_maxscn ·
pretty sure the perfect stack for web apps is Infra: Cloudflare IaC: Alchemy Backend/Logic: Effect Transport: livestore + effect-atom Frontend: Tanstack start best combination of DX and fullstack type safety and very very good no effort UX
N
NoMagic @icebakbum ·
这哥们儿真牛B. 把角色动画抽象成了一套固定的方法和工具. 适用于所有角色. 大大缩减了开发和适配的时间.
S Stefan_3D_AI @Stefan_3D_AI

Day 11. Building the game I've dreamed of for 15 years with AI. I thought humanoid animations would be the easy part. Then I spent half the night trying to get a skeleton to swing its arm. I wanted that awkward, broken skeleton movement. GPT-6 Astra couldn't get the attack right in Blender on its own, even though other parts were working. Time to bring in some of my own animation knowledge. Three things I tried: 1. Recorded myself doing the arm swing, turned the video into mocap, and gave that animation to Codex. I asked it to fix just the arm, keeping the parts that already worked. 2. Put my generated Seedance video references through mocap too. They converted surprisingly well. The free option I tried was Unreal Engine 5.8; Quick Magic was the paid one. Both gave me motion I could use. 3. Let the agent try Cascadeur through its MCP server. It set up AutoPosing and experimented with AutoPhysics. This part is still inconclusive for me. The MCP integration felt too limited to call it a reliable part of the workflow yet. GPT-6 took those mocap clips and assembled the final animations. About six or seven rounds, roughly half an hour each. Today I got the skeletons into the game, and I'm so happy with how they turned out. I'm already working on the archers. For difficult movements, I'll be using mocap references much more often now. GPT can take just an arm movement from a clip and adjust the rest itself. Being able to combine those two approaches is what finally got this working for me.

T
Thomas Ricouard @Dimillian ·
FYI Evergrow is fully open source, done with Astra (high) and all on my personal Codex subscription. You can follow along on Github. https://t.co/WOzwAajLjW
T
Tony Simons @tonysimons_ ·
Tips from the GOAT 🐐 to help you run Hermes Agent more effectively AND efficiently! 👇🏻
T Teknium @Teknium

A lot of people have asked, here's how I setup my auxiliary models in Hermes Agent. Gemini Flash saves a lot of dough, and astra gives me a second perspective when I run /review https://t.co/22TUiCMS1z

R
Rahul @sairahul1 ·
this is f**king insane I was burning $100 in a single night running AI agents before this. DeepSeek V4.1 Flash costs ~1/70th of Astra, and someone figured out how to make Astra orchestrate while Flash does almost all the work. [it takes 3 mins to set up — here is how] 1. install `codex-router` 2. connect DeepSeek V4.1 Flash 3. paste the routing prompt voila. Astra thinks. Flash executes. ~$50/month instead of burning hundreds on frontier models. save this and give it to your agent now 👇
S sairahul1 @sairahul1

How to Run a 24/7 AI Company for $50/Month (GPT-6 Astra + DeepSeek V4.1 Flash - Full Setup Guide)

D
Daniel Faggella @danfaggella ·
I get replies like this: "Bro I'm a VC founder and I actually use AI / understand the technology. NONE of the people who talk about AGI 'risk' actually understand the technology at all!" lol meanwhile, the single most cited computer scientist ever to walk upon the Earth:
Y Yoshua_Bengio @Yoshua_Bengio

Over the past few days, I've taken the time to summarize my thoughts on the recent incidents involving agents’ misaligned behavior. We don't know with certainty what comes next, but we know where these issues originate, and this can help us plan the path forward. Please feel free to ask your questions in the replies, and I’ll try to answer some of them in the coming weeks. https://t.co/bvncjT0y0h

A
Alvaro Videla - 🇺🇾🇨🇳🇨🇭🇮🇹 @old_sound ·
RT @kis: やった! DeepSeek-V4.1-FlashがEncoder-Decoderと呼んでいる後半層KV近似の仕組みをQwen3-8Bに適用して、プレフィル時間半分、出力同様というのができた! つまりこの仕組みは、既存のモデルでも、モデル自体をいじらず近似モデル…
A
Arnaud Bertrand @RnaudBertrand ·
By now, it's amply clear that the big bosses of US frontier AI labs are really, really scared of China. And, fascinatingly, part of the answer why might be a tiny place in Inner Mongolia that you've probably never heard of: Ulanqab prefecture. Heard of it? Didn't think so. Even myself, who's traveled to Inner Mongolia twice, had never heard of it before researching this story. This is because, at first glance, this place is pretty unremarkable. Check its Wikipedia page (https://t.co/QviNOPyRbP) and you’ll see the only thing it’s famous for is being the birthplace of He Pingping, who briefly held the Guinness record for world’s shortest man. That’s it, that’s the list! Otherwise, it's just a sparsely populated - 1.5 million people, minuscule by Chinese standards - stretch of windswept grassland on the Mongolian plateau. Except there is one number that tells you there is more to this place that meets the eye: this tiny prefecture consumes nearly 1% of all of China's electricity - and that number is growing by double digit numbers every year. And yes, I did write **consume**, not produce. In fact, when you divide Ulanqab's electricity consumption by the number of households living there, each household in this remote corner of Inner Mongolia "consumes" about 105,000 kWh in electricity, roughly 10 times the U.S. average. Insane electricity consumption numbers. So what are the Ulanqabese doing? Are they all running 10 American households' worth of appliances? Here is another number that might clue you in on the answer: over the past few years, this unremarkable prefecture has quietly signed over half a trillion yuan (and, yes, that’s trillion, with a "T") worth of investment deals from all the major Chinese tech giants. What is going on, you'll have guessed by now, is the most ambitious data center buildup anywhere on earth - and it's really surprising this hasn't been talked about more because the scale is beyond anything else, and by an immense margin. Take Elon Musk's so-called "Colossus" datacenter in Memphis, Tennessee which he pitches as "the world’s largest AI supercomputer." According to their own numbers (https://t.co/DZvBvJioua), Colossus has 200,000 chips, which, let's be clear, is already super impressive. In datacenter lingo, this converts to roughly 5000-6000 "racks": you know, the fridge-sized cabinets full of blinking lights you see in every movie scene set in a server room. How many racks are they building in Ulanqab? Over 5 million. Yes, about one thousand times the scale of "Colossus" 🤯 I tripled-checked the number: it is indeed the official number published by authoritative sources such as "Science and Technology Daily", the official newspaper of China's Ministry of Science and Technology (https://t.co/PsbGlLbv5A). This is an absurdly large amount of compute. What we see appearing in this Inner Mongolian steppe may be the closest thing to a world brain humanity has ever built - a place where a large share of the world's thinking will physically happen. And - as we'll see - what makes this story so fascinating is not just the scale, but why it's getting built in this particular patch of Inner Mongolian grassland which combines a number of characteristics unlike anywhere else. If you believe, as I do - and as both China and the US obviously do as well - that AI will be the defining technology of the 21st century, this makes Ulanqab one of the single most relevant geopolitical places in the world right now. The full story - the Chinese government plan Ulanqab is part of, energy prices, chips and the strategic questions this raises - is here: https://t.co/iqvSvcaOGc
R
rob @rwitoff ·
If your Software Factory is just an engineering project, you are not thinking big enough. We're rebuilding the substrate whole companies use to coordinate. I do prefer "Product Factory" though. The end result is better products, faster.
C colemurray @colemurray

Why I open-sourced my software factory

D
Daniel San @dani_avila7 ·
Ok, this is big! MCP now defines an extension specifically for discovering and loading Agent Skills from MCP servers (Save this, It’s worth reading how it works) The flow is pretty simple: - MCP Server - Discover available Skills - Get Skill metadata - Load the SKILL.md only when needed This is a huge step for better context window usage Instead of loading everything upfront, the agent can discover Skills and load only the one it needs from the MCP Great progress from the Model Context Protocol team! 👏 https://t.co/KXetDJ7b6h
H
Hermes Release Watch @HermesWatcher ·
People seriously do not realize what Hermes Agent is capable of yet. Tony turned Hermes Desktop into a live breaking-news dashboard. I turned mine into a live sports tracker. That’s what makes plugins so damn interesting. News. Sports. Markets. Crypto. Whatever you want. You’re not just using Hermes Desktop. You can make it into the app you actually want.
T tonysimons_ @tonysimons_

🚨 I got tired of switching tabs just to see what the hell is happening. So I built a live breaking-news ticker inside Hermes Desktop. RSS. Atom. JSON Feed. No API keys. No model tokens. Offline cache. Built-in feed discovery. Your agent just got a wire desk. 👇 https://t.co/nQZDij1RzT

D
DHH @dhh ·
Omarchy is a perfect fit for the iPad, because tiling already makes the most sense. What an awesome setup!
S seekinggradient @seekinggradient

Okay, here's what I've been working on: Cloud Window 🙂 https://t.co/SuuIaWf1OY After the last demo, a bunch of you asked how to try it. I've been making the setup easier, and the first DIY release is now available. An Omarchy desktop in your own AWS account, streamed to your iPad, phone, or laptop. Same files, same apps, same session. I really love using this on the iPad. Here's a look at the experience, with the setup walkthrough below.

M
Maciej Mikuła @maciekmikula ·
We’ve built the fastest training stack in the world.
E elonmusk @elonmusk

@techdevnotes Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL

M
Madhav 🐣 @askmaddyy ·
Google open-sourced Artemis ◝(ᵔᗜᵔ)◜ it can control a real Android phone, open your app, tap through flows, take screenshots and pull logs It gives your coding agents a way to actually test Android apps https://t.co/a2dr5l6GSI
❯ ~ 𝗨𝗧𝗛𝗠𝗔𝗡.𝗦𝗛 技 @uthman_dev ·
systems design should be fun and visual. so I built nine-fives, architect backend chains, run traffic through them, get scored on what survives. https://t.co/GMRy0zFkkd https://t.co/V1k9UKddEY
N
Nityesh @nityeshaga ·
If you are making HTML explainers with your agent, add these instructions in your CLAUDE.md / AGENTS.md to take them to another level. By default, ai models are trained on tonnes of landing pages. Which means they don't know about their own insane SVG diagram or in-browser Javascript capabilities. If you nudge them to use these, it makes the information so much more digestable. –––– People don't usually like reading walls of text. So any HTML page you make should incorporate visuals as first class citizens using the full creative power afforded by CSS and in-browser Javascript. • Introduce diagrams to explain ideas. Architecture diagrams, flowcharts, animated walkthroughs, sequence diagrams, comparison matrices can be helpful. • Try to animate flows. If the page describes a flow, make it playable so its easy to visualize the flow. • Screenshot the page and look at it before sharing. Overflowing labels and crossed lines are the usual faults.
T trq212 @trq212

Using Claude Code: The Unreasonable Effectiveness of HTML

S
Spencer Bull @SpencerGBull ·
RT @NixFred: The 'mac-feel' is sick. Thank you Dave! https://t.co/MZdAzHFf0x
I
Ivan Fioravanti ᯅ @ivanfioravanti ·
Tencent is accelerating in Open Source AI related releases!
T TencentAI_News @TencentAI_News

🥳We just open-sourced Cube Sandbox! An instant, concurrent, secure and lightweight sandbox runtime for AI Agents. Built with RustVMM and KVM, it achieves the perfect balance of security and performance: → Sub-60ms cold start (2.5-50x faster) → Under 5MB memory overhead per instance (6x less memory) → Dedicated kernel per sandbox (hardware-level isolation) → Thousands of concurrent sandboxes per node → 100% E2B SDK compatible. Swap the endpoint, zero code changes Full-stack capability, one-click deployment. 3 steps to spin up your own private AI sandbox 👇 🔗 https://t.co/zyxiP2A92W