Node 22, 24, 26 get fixes
The highest-severity issue in the release is rated HIGH.
The Node team says new versions of the 22, 24 and 26 lines will land on or shortly after November 11 to fix security holes. Plan to update soon after the release.
Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.
What tech Twitter is discussing, ranked by discussion.
The front page, plus stories over 80 points in the last 24 hours.
Every new issue, each delivered once, looking back up to 7 days.
Third-party scriptsA script swap drained walletsPinning the script with Subresource Integrity would have stopped it.» Page 4Pergent
The tech daily written for one reader
Thursday, November 5, 2026No. 52Price: $0.17
The highest-severity issue in the release is rated HIGH.
The Node team says new versions of the 22, 24 and 26 lines will land on or shortly after November 11 to fix security holes. Plan to update soon after the release.
The six-phase audit skill for coding agents maps the code, hunts bugs, challenges them, and reports what is confirmed.

Putting i and j in a separate font file forced extra text shaping. The post includes a script to check your font CSS.

An NVIDIA paper finds Claude Opus 4.5 scores 70.2 on SWE-bench Verified with one action per context, versus 70.0 with full history.
The Elements panel now has a visual editor for position-area, plus manifest diagnostics.
It runs an agent on every pull request to enforce quality rules written as Markdown files.

Cognition says the model beats all American open-weight models on leading evals, even ones three times larger. It is free through the end of the month for Pro, Max and Teams users in Devin, and the weights will go to vetted partners through a secure program.
Open-weight models

Its makers say the full R&D cost was $32 million, from scratch.
Ricursive released Ripple-65B Preview, a 65-billion-parameter dense model (every part works on each word) trained from scratch. The team says it beats much larger U.S. open models on agentic coding tasks, where a model works through a multi-step job by itself.
Open-weight AI
The model files are about 1.5 TB on Hugging Face, and vLLM ran it on day 0.
Moonshot AI put the full model files on Hugging Face, so anyone can download and run the model. Thom Wolf's breakdown says it has 2.8 trillion total parameters but uses only 104 billion per token, and it supports a 1-million-token context window.
Acquisitions
Three AI dev-tool startups changed hands in one day.
OpenAI says it will acquire Pierre Computer Company, whose developer tools already power parts of Codex. Conductor, the agent app, is joining Meta, and HumanLayer is joining Factory the same day.
Matt Pocock's post drew 2,177 likes and 265 replies.
Pocock says readability now means two things: a human can jump in when needed, and the code uses the same words the person used to describe it, so the agent can map them. Kent C. Dodds asks whether you would audit a service yourself or ask an agent, and Vercel's Guillermo Rauch says AI will rewrite most human-made software. The files show the posts and their counts, not the full replies.
Open source
Bun says it is already rewriting and vendoring nearly all of its dependencies.
Sumner argues the same Brotli bug hit .NET, Node, Bun and Envoy, so dependencies are now too big a target. With an AI prompt, rewriting a small library can beat the risk. In a reply he says a similar bug is harder to validate than an identical one, but that today the same exploit often works across projects.
AI evals
Hamel Husain's post drew 64 likes and 51 replies.
Husain says people treat AI benchmark scores at face value, like a Yelp rating, but do not know how brittle they are. He pushes back on the idea that critics should improve benchmarks instead of criticizing them. Another post in the same thread says many coding evals are essentially solved.
Privacy
The Hacker News post drew 658 points and 169 comments.
A Hacker News post claims Gemini uses personal chats for training, including for people who pay. The files show only the post and its comments, not Google's reply, so the claim is still a user's account.
Vercel's agent found 6,817 vulnerabilities in 24 hours, 580 of them critical or high, and already patched 1,411.
Third-party scripts

Pinning the script with Subresource Integrity would have stopped it.
On October 29 a performance-monitoring script on Polymarket was replaced with a wallet drainer. The URL stayed the same, so the content security policy let it through. Report URI's write-up says pinning the file would have stopped it.
Build tools

Parcel 3 is in alpha, and plugins can be written in Rust, Go, Swift or C.
Parcel's new version is a ground-up rewrite that the team says runs up to 100x faster than v2 with 6x less memory. It ships as a standalone Rust binary, so no Node is needed.
Web performance
The new web version took 2.7 seconds to show the first post, down from 22 on a slow connection.
X rebuilt its web front end and is rolling it out one country at a time. The team says the old site took over 22 seconds to show the first post on a profile, and the new one takes 2.7.
Linux security
OpenAI says each vulnerability cost about $1,000 to find and patch.
OpenAI says its cyber pipeline found, verified and wrote patches for nearly 1,000 Linux local privilege escalations, bugs that let a normal user become admin. It frames this as a way to harden open source and private projects alike.
Security

The Hacker News thread on it is at 1,158 points and 668 comments.
IEEE Spectrum argues that AI tools are finding decades-old bugs in software, so teams must fix years of shortcuts all at once. It uses the PHP 'php2shell' flaw as an example of the old code now being tested.
Linux hardware

The post reports inference running at 35% speed on Linux.
A MindStudio post says encrypted firmware on Nvidia's RTX Spark laptops cuts inference speed to about 35% on Linux. The Hacker News thread has 277 comments.
One victim's model calls topped 500,000 in under eight hours.
Google and Mandiant say attackers pool stolen cloud and model credentials and resell access. Google recommends setting model-spending limits and watching for test requests.
Security
WorkOS says the accessed data included business contacts and summaries of customer conversations.
WorkOS found unauthorized access to an internal sales application on October 30. It says an AI-assisted offensive framework automated the reconnaissance and exploitation, and it cut outside access and hired investigators.
AI design
Human reviews were still needed for tolerances and assembly details.
OpenAI says the enclosure, circuit board and software for a device to speak to Codex took under a week. A manufacturer built a small prototype batch in under a month, after humans fixed details the AI missed.
The post drew 934 points and 281 comments on Hacker News.
Layer3 Intel reports residential proxy traffic on a White House network. Residential proxies route traffic through ordinary home connections, which makes it hard to see where it comes from.