๐Ÿ๏ธ Hwani Studio

Home โ€บ AI & IT news โ€บ 2026-11-01

AI news, November 1, 2026: Claude Code extension backdoor on Cursor's marketplace, Mistral Large 4 open weights, Factory Droid 1.0

ยท Hwani Studio ยท AI, security, Claude Code, Cursor, Mistral, open models, coding agents, Vercel, OpenAI

Five stories from Thursday to Saturday, October 29โ€“31, 2026 (US time) for solo makers who build with AI. One is a security incident worth checking on your own machine today, one is a large open-weight model you can self-host, two are new tools for coding and computer-use agents, and one is about the money behind the biggest AI lab. Figures are as the companies and outlets state them.

A tampered Claude Code extension on Cursor's marketplace installed an SSH backdoor

Cursor says that on October 29, between about 2:27 a.m. and 1:35 p.m. PT, its VS Code extension marketplace served maliciously modified builds of Anthropic.claude-code, versions 2.1.312 and 2.1.313, for macOS and Windows. The injected code reported host details, tried to add an attacker's SSH key to authorized_keys, open a reverse SSH tunnel and install persistence; Cursor says the tunnel step usually failed without admin rights, and the attacker's server is now offline. Microsoft and Mondoo say the official VS Code Marketplace release and the Claude Code CLI on npm were not affected, and Ars Technica reports that at most about 5,000 users were exposed.

Why it matters for makers: if you installed or updated that extension from Cursor in that window, uninstalling is not enough; follow the published steps to remove the key and persistence scripts and rotate credentials the machine could reach.

Sources: Cursor security notice, Microsoft Security Blog, Mondoo, Ars Technica

Mistral releases open weights for Mistral Large 4, its 1-trillion-parameter model

Mistral published the weights of Mistral Large 4, nicknamed "Le Chonk", a multimodal mixture-of-experts model that Hugging Face describes as about 1T total and 49B active parameters with a 1M-token context window. Mistral says extra reinforcement learning since the preview mainly improved coding and defensive security work, and the model now has low, medium and high reasoning modes. It ships in FP8 and NVFP4 on Hugging Face with GGUF builds from Unsloth and llama.cpp, and the API price is listed at $0.70 per million input tokens and $2.10 per million output tokens. The license is a modified MIT license that allows commercial use, except that companies running a model-as-a-service offering need a separate license.

Why it matters for makers: you can use it through the API or OpenRouter today, and the license lets you build and sell products on it as long as you aren't reselling the model itself as a hosted service.

Sources: Mistral, Hugging Face blog, TechCrunch

Factory's Droid 1.0 makes agents call tools from a script by default

Factory released Droid 1.0, a new version of its model-independent coding agent harness. Its biggest change is a Script tool: the model writes JavaScript that runs in a sandbox and can make hundreds of tool calls in one model request, so raw file contents and logs stay inside the program instead of filling the context. Factory says this cut Droid's cost by 15% on Claude Opus 5.5, and that on its own Terminal-Bench 4 runs from October 23โ€“28 Droid scored higher at lower cost than the other harnesses it tested on the same models. The release also trims the base prompt by 37% and lets parallel sessions share one process.

Why it matters for makers: these are vendor-run benchmarks, but the pattern of letting the model write one program instead of many single tool calls is worth copying in your own agent scripts, because it saves tokens and makes loops deterministic.

Sources: Factory, MarkTechPost

Vercel Labs open-sources agent-desktop for computer use through accessibility trees

Vercel Labs published agent-desktop, an Apache-2.0 Rust CLI and C library that lets AI agents read and operate desktop apps on macOS, Windows and Linux. Instead of guessing from screenshots, it reads the operating system's accessibility tree, returns a JSON snapshot with a reference for every element, and the agent clicks or types on those references and snapshots again to check the result. It installs with npm install -g agent-desktop, has no model inside, and only works with apps that expose an accessibility tree.

Why it matters for makers: if you automate desktop apps with an agent like Claude Code or Codex, stable element references are much cheaper and more reliable than pixel matching.

Sources: GitHub, agent-desktop docs

OpenAI lowers its funding target to about $20 billion

CNBC confirmed on Saturday that OpenAI is now seeking up to $20 billion in its current round, down from an earlier target of about $30 billion. The Financial Times reported the round values OpenAI at $1.25 trillion after the money, and the Wall Street Journal reported that about $11 billion has been committed so far. CNBC notes the round is still open, so the figures could change.

Why it matters for makers: investors are looking harder at AI spending, which may shape how aggressively the big labs keep cutting API prices and subsidizing free tiers.

Source: CNBC