Vite dev server flaw
Fixed in 8.3.7, 8.2.5, 8.1.7, 7.3.8 and 6.4.6.
Anyone running the Vite dev server with --host or server.host, so it is reachable on a network, can be hit by remote code execution. Update right away if you expose your dev server.
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.
npm now scans packages on publish» Page 4Pergent
The tech daily written for one reader
Saturday, November 14, 2026No. 61Price: $0.05
Fixed in 8.3.7, 8.2.5, 8.1.7, 7.3.8 and 6.4.6.
Anyone running the Vite dev server with --host or server.host, so it is reachable on a network, can be hit by remote code execution. Update right away if you expose your dev server.
Anthropic redesigned Projects so several AI workers can run at once. The Hacker News thread on it has 76 comments.
Anthropic status
The status incident covers claude.ai, Claude Code and the developer platform.
Anthropic logged elevated errors on claude.ai, Claude Code and its platform console, and the HN thread is 384 comments long. Gergely Orosz says Claude was down for his task, so he moved to Grok Bot for it, which is his case for not relying on one AI subscription.
A set of skills that point agents at modern, Baseline-aware HTML and CSS, plus a list of the top 10 things AI gets wrong in web dev.
Themes are built on the mods system, with animated backdrops and custom colors, and you can ask Claude to generate one.
Claude now supports more of the MCP Apps extension, which lets tools show interfaces.

Vercel says steps start up to 25 times faster, with less overhead.
Version 5 encrypts every payload by default, adds regional failover, and cuts time to first step and time between steps. Guillermo Rauch says the reliability has been tested at large scale. The part other tools may use most is durable-harness, a reusable durability log that any agent setup can plug into.
Cognition
Apps run on Devin's cloud computer and can be shared privately.
Devin Apps are meant for internal tools, preview deployments and extensions to Devin itself. Kent C. Dodds says his own Kody platform already does this, and notes Cursor, ChatGPT and Claude each have a similar feature.
Local AI

The first fully stable release promises support for every hardware type and model.
llama.cpp is the open-source library that runs AI models on your own laptop, phone or workstation, with no cloud needed. Version 1.0 is the first major stable release, and the team says it will keep covering every platform and model. Hugging Face's Clement Delangue calls it a way to own your intelligence. A matching ggml v1.0.0 also shipped, the math library underneath it.
OpenAI
The tools aim to help small teams keep up with the flood of AI-written code.
Daybreak Red Team is a group of security agents that check whether code flaws can really be exploited. Daybreak Defense Factory helps fix what Red Team finds, building on Codex Security. Red Team is in early access and Defense Factory is in the Daybreak plugin for Codex; no pricing was shared.
Anthropic
Investor materials put the work at about three days, against six to twelve months the usual way.
Anthropic says its Mythos model wrote or adapted nearly all the chip-specific code for one Claude model's prompt processing on AMD MI355X GPUs. The figures come from investor materials and are not independently verified. The result hints at lower switching costs between chip makers, but it does not prove every Claude workload can move freely.
OpenAI
Agents pay without ever seeing your card details, and every purchase needs your approval.
ChatGPT Wallet rolls out today to dots, for Pro subscribers in the US. Each purchase gets a security check, and it works with American Express, Mastercard and Visa.
His video has more than 250 comments on Hacker News.
The video asks a simple question: if AI has solved coding, why does software still look the same? It is a sharp version of the argument about whether AI tools have made real products better, not just faster to write.

npm validates publishes with a security scan that can add a little wait; Kent C. Dodds says it is worth it.

It starts in 2 ms, and npm, pnpm, yarn and bun link straight to a native binary, with no Node wrapper.
Security

About 100,000 workspaces may have had their secret variables decrypted.
Railway says attackers got a copy of its code, then used AI agents to probe several routes in parallel until they reached a system holding encrypted customer variables. The cloud host estimates unauthorized use of the stolen keys at about $740,000. The 100,000 figure is an upper bound, not a count of confirmed victims.
Microsoft 365
The fix took about 20 hours because it had to move forward, not roll back.
Follows up on the Nov 12 Microsoft 365 and Outlook outage. Gergely Orosz says Microsoft engineers told him the cause was a tangle of networking problems, not DNS as many guessed. The fix meant moving pieces from a legacy Exchange system over to Azure, plus registry patches, rack reboots and traffic routing to handle a thundering herd. The postmortem is expected to be public.
Software Mansion's Web Audio API for React Native adds spatial audio, and FFmpeg is now opt-in.
Paul Irish's tool records a Chrome tab with a single npx command, handy for PR clips.

React added docs on preserving form values after an Action, plus a reference page for requestFormReset.
OpenAI
GPT-6 Astra answered safely on 91.5% of fixed cyber jailbreaks, versus 59% for GPT-5.6 Sol.
OpenAI trained models on prompts from an AI attacker and reports stronger resistance to fixed and adaptive jailbreaks. The study covered cyber, bio and violence. These are test results on evaluated prompts, not the full safeguards used in production.
Sebastian Raschka lists more than a dozen launches since late September.
A decision model is a small model that scores or picks among options instead of writing text. Launches came from Together AI, Alibaba, Liquid AI, Databricks, AWS, Cloudflare, Perplexity, OpenAI, Snowflake, Microsoft, IBM and more. Perplexity says it now handles a billion decision requests a month on OpenRouter.
Anthropic
The guide covers how to design, run and report safety evals.
Safety evaluations help decide whether a model is ready to ship, which safeguards to build, and where to focus training. Anthropic's Engineering Blog gives 10 recommendations for doing them well.
World models
The write-up drew 40 comments on Hacker News.
Haddy Alchaer shares what came out of 600 experiments on JEPA world models. A Northwestern talk this week argued JEPAs are the next step for learning from any data.
Ami Luttwak, Wiz co-founder, says attacks that took weeks now take minutes.
Luttwak says companies should keep attacking their own systems with the latest AI models to stay safe. He says attackers can get in and take secrets and AI keys in about 20 minutes, so startups need automated testing and patching from day one.
Reverse engineering
The post hit 735 points and 211 comments on Hacker News.
A developer reverse engineered a Coway air purifier so it no longer needs the company's cloud service. It is a clear walk through how to undo cloud lock-in on a home device.