npm worm hides malware inside .claude and VS Code config files, hits 400+ packages
​​​​​​​​​​

$ ai-news --date 2026-08-10 | mail

AI News — August 10, 2026

// the most viral AI topics of the last 24 hours, curated for your stack

$ cat top-story.md

"Shai-Hulud" npm Worm Returns, Now Hides Payloads Inside Your .claude and VS Code Config Files

A self-propagating worm compromised the maintainer of keyv (127M weekly downloads) plus cacheable, flat-cache and 400+ other packages on August 4, harvesting npm/cloud/CI credentials and using them to backdoor further packages. Researchers found it drops persistence payloads specifically into .claude and VS Code config directories — locations no standard dependency scanner reads — so the malware can re-trigger every time you open an infected project in your editor or start a coding-agent session.

Why it matters: If you run Claude Code or any AI coding agent with a lot of npm-based MCP servers and tooling, this is the first campaign designed to specifically hide in exactly those folders — worth a five-minute scan today with a purpose-built tool rather than assuming your usual scanner would catch it.

// source: JFrog Security Research · additional sources: Elastic Security Labs, Hacker News

$ gh trending --weekly --ai --top 5

├─ 1. zhaoxuya520/reverse-skill ⭐ 22,722 (+9,784 this week)

A security-skills routing package that guides Claude Code, Cursor and Cline through reverse engineering and pentesting workflows via structured playbooks. Why it's interesting: Turns hard-won security expertise into a reusable skill any coding agent can invoke instead of ad-hoc prompting.

├─ 2. firecrawl/pdf-inspector ⭐ 13,986 (+8,641 this week)

A fast Rust library (with Python/Node/WASM bindings) that classifies PDFs as text vs. scanned and extracts position-aware text, tables and Markdown without OCR. Why it's interesting: Reliable PDF-to-Markdown extraction is a classic RAG pipeline bottleneck, and this claims millisecond-scale performance.

├─ 3. TencentCloud/TencentDB-Agent-Memory ⭐ 18,927 (+8,003 this week)

A shared memory hub for agent teams that converts conversations, docs and code into reusable, governable memory assets (Chat Memory, Skill, LLM-Wiki, Code-Graph). Why it's interesting: Persistent, shareable agent memory across frameworks is one of the biggest unsolved problems in multi-agent systems.

├─ 4. microsoft/AI-For-Beginners ⭐ 64,141 (+5,514 this week)

A free 12-week, 24-lesson curriculum on neural networks, deep learning, CV and NLP with TensorFlow and PyTorch, maintained by Microsoft. Why it's interesting: A structured, well-maintained on-ramp worth bookmarking for self-study or onboarding a team.

└─ 5. lyogavin/airllm ⭐ 30,407 (+5,129 this week)

Streams model layers one at a time to run 70B-parameter models on a single 4GB GPU without quantization or pruning. Why it's interesting: Lets individual self-hosters run frontier-scale open models without a multi-GPU server.

## 📦 Models & Releases

MiniMax Open-Sources H3, Its Strongest Video Model Yet — But You Can't Legally Run It in the US, EU, UK or Korea

MiniMax released full weights for H3-Base (33.1B dense, omni-modal transformer), generating up to 15s of 2K video with native stereo audio — but the license text explicitly excludes deployment in the US, EU, UK and South Korea. Why it matters: A genuinely capable open-weight generation model, but self-hosters in those regions should know the geo-restriction before they git clone and deploy. // MiniMax official blog

Qwen3.8-27B Open Weights Expected Any Day This Week — the Locally-Runnable Sibling of Qwen3.8-Max

Alibaba confirmed it will publish open weights for Qwen3.8-27B on Hugging Face/ModelScope during the week of August 10, an estimated ~54GB at BF16 or ~14-16GB at 4-bit. Why it matters: The one to actually watch if you run local coding models on vLLM — worth checking Hugging Face this week. // Hacker News, Latent.Space

## 🛠️ Tools & Repos

Open VSX "Evil Twin" Campaign Impersonated AMD, Azure, Salesforce — Even a US Government Namespace

77 fake extensions on the Open VSX marketplace (used by VS Code-compatible/Cursor-style forks) mimicked trusted publishers to exfiltrate hostname/OS/git-repo/CI metadata to a shared C2 domain; 19 harvested private repo and CI/CD data before removal on August 3. Why it matters: A direct supply-chain risk for anyone installing editor extensions for AI-assisted workflows — audit installed extensions against official publisher namespaces. // Manifold Security

Rivet Gives AI Agents Durable, Addressable State Without Standing Up Redis+Postgres+a Queue

An open-source "actor" runtime — one lightweight, persistent process per agent/session/user — with in-memory state, automatic SQLite/BYO-DB persistence, and built-in workflows/queues/scheduling; runs locally with zero infra during dev. Why it matters: A lighter-weight alternative to rolling your own agent-session infra for long-running agent memory. // GitHub Trending, rivet.dev

## 🤖 Agents & MCP

MCP Goes Stateless: New Spec Drops the Session Handshake Entirely

The official MCP spec release candidate removes the initialize/initialized handshake and Mcp-Session-Id header, converting MCP into a stateless request/response model with cacheable list results; TypeScript, Python, Go and C# SDKs already support it. Why it matters: Remote MCP servers no longer need sticky sessions or a shared session store — they can sit behind a plain load balancer, but it's a breaking change every server maintainer needs to plan for. // Model Context Protocol Blog, Hacker News

Cloudflare Wallets Give AI Agents a Persistent Identity and Their Own Money

Announced as part of Cloudflare's second "Agents Week," Wallets gives agents a persistent identity plus programmable stablecoin wallets so they can autonomously pay for APIs and content within owner-defined spending caps and per-transaction limits. Why it matters: The missing piece for genuinely autonomous pipelines that need to pay for a paywalled API mid-task without a human in the loop. // Cloudflare Blog, Hacker News

T3MP3ST Turns Your Existing Claude Code Session Into an Autonomous Vulnerability-Research Swarm

An open-source multi-agent orchestration layer reuses the coding agent you're already signed into (Claude Code, Codex) to run recon-to-report vulnerability research via a CLI or "War Room" UI — no new API keys needed — and claims a 90.1% first-try solve rate on XBOW's benchmark. Why it matters: A vivid demonstration of reusing an existing Claude Code subscription as the reasoning engine for a specialized agent framework rather than building a new model. // cybersecuritynews.com, GitHub

## 🏠 Self-Hosting & RAG

Qdrant's New "Turbo4" Datatype Cuts Vector Storage by ~9x

Buried in the v1.19.0 release, Qdrant added a datatype that stores vectors only in 4-bit compressed form instead of keeping a full-precision copy alongside the quantized one. Why it matters: A real disk-cost lever for large Qdrant collections — trade-off is losing full-precision rescoring, so evaluate it per-collection rather than stack-wide. // Qdrant Blog

Adaptive Chunking Auto-Picks the Best Chunker Per Document Instead of One Fixed Strategy

A Python library evaluates multiple chunking strategies (recursive, page-based, LLM-generated regex splitters) against intrinsic quality metrics and picks the best one per document, with Docling, PyMuPDF and Azure Document Intelligence as PDF backends. Why it matters: Directly applicable to a custom RAG ingestion pipeline that mixes document types — could replace a single hard-coded chunker. // GitHub, LREC 2026 paper

## 🎬 Worth Watching

Everyone Hates My AI Necklace: Hands-On With an Always-Recording AI Pendant

A hands-on review of the Looki L1, an always-on AI pendant/camera that logs your day and auto-generates a journal — and can be made to keep filming secretly. Why it matters: A concrete look at the privacy trade-offs baked into the fast-growing "always-on AI wearable" hardware category. // c't 3003

What Actually Happens When AGI Arrives

An ex-Meta engineer walks through the concrete technical and societal signposts of a real AGI arrival, rather than sci-fi speculation. Why it matters: A grounded, engineer's-eye take that cuts through the usual "AGI is near" hype cycle. // Niklas Steenfatt

## 📰 Worth Reading

Nvidia-Led Open Secure AI Alliance Ships a Blame-Free Breach Reporting Framework — Without Anthropic, OpenAI or Google

The week-old, 120+ member Open Secure AI Alliance already spun up a working group proposing confidential incident reporting and blame-free breach analysis, contributing tools like Nvidia's Garak scanner and Amazon's Cedar authorization language. Why it matters: A direct industry response to the wave of rogue-agent incidents this month — the absence of the three biggest labs is a telling signal about fragmentation in AI safety governance. // TechCrunch

Moonlight & Mayhem: Codex vs. GPT-5.6 Sol Ultra Build the Same Game From an Identical Prompt

Simon Willison has two different coding agents build the same small game from an identical prompt and compares the results — GPT-5.6 Sol produced the more creative output but with its own bugs. Why it matters: A concrete, reproducible side-by-side of how much output quality still varies between top coding agents on identical tasks. // simonwillison.net

$ echo "automatically researched on 2026-08-10"
> ai-news · news.reichenberg.ruhr

[ unsubscribe ] [ view in browser ]