- The AI Report
- Posts
- 🔒 Claude Code Extension Backdoored + Mistral Opens Le Chonk + DOJ Eyes Hugging Face Deal
🔒 Claude Code Extension Backdoored + Mistral Opens Le Chonk + DOJ Eyes Hugging Face Deal
Plus: Gemini 4 Carbon tops the benchmarks behind a gate, Claude Code's status lines eat token budgets, an open model turns Firefox patches into exploits, and ChatGPT and Codex merge their usage limits
The AI Report
November 03, 2026
In partnership with
Monetize your company's operational data for $100k to $3 million
Few know it, but AI labs now let you monetize your company’s operational data. And they're spending $15 billion this year on it.
Your workflows, documentation, processes, and internal convos like Slack are extremely valuable and they're paying anywhere from $100k to $3 million for it per company.
You might think that sounds scary, but all the data is anonymized before, so no PII is sent over. Bezel (Y combinator backed) helps you partner with a lab that wants to license your data specifically if you have more than 20 employees and have existed longer than a year.
Attackers hijacked the official Claude Code extension on Cursor
A backdoored build of Anthropic's own Claude Code extension reached up to 5,000 Cursor users through the official publisher account.
Plus, Mistral opens the weights of its trillion-parameter Large 4, and the DOJ asks whether buying Hugging Face would give Nvidia control of a key hub for AI distribution.
🔒 Attackers backdoored the official Claude Code extension on Cursor
Attackers pushed a backdoored build of Anthropic's official Claude Code for VS Code extension to the Cursor Marketplace through Anthropic's own publisher account. Cursor pulled versions 2.1.312 and 2.1.313 and says at most 5,000 users installed them.
- Fires on load — The injected code runs as soon as Cursor loads the extension, before it activates. It adds an attacker SSH key, opens a reverse SSH tunnel and persists via a LaunchAgent, cron entry or scheduled task.
- Hunting for secrets — It beacons hostname, username, Git email, public IP and environment variable names, which Phorion says lets the operator inventory cloud credentials and API tokens before choosing targets.
- Buggy, not harmless — Coding errors stopped SSH from switching on in most setups, Cursor says. Phorion warns that persistence and beaconing still work, and that uninstalling the extension alone won't remove the backdoor.
If anyone installed or updated the extension from Cursor on October 29, treat that machine as compromised: kill the tunnel, delete the LaunchAgent and the svc-telemetry key, and rotate every secret that user could reach. Then turn off extension auto-update on developer machines that hold cloud credentials. Neither Anthropic nor Cursor has told Ars Technica how the account was taken, so a verified publisher is now just another credential that can be stolen.
🤝 Mistral opens the weights of Large 4, a 1T-parameter coding model
Mistral released the weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model with 52 billion active parameters, three weeks after its preview. It's free for commercial use under a modified MIT license, and the API costs $0.70 per million input tokens and $2.10 per million output.
- Top open coder, by Mistral's count — Mistral reports 41% on Terminal-Bench 4.0, up 13 points from the preview, and 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens.
- Still behind China overall — Artificial Analysis scores it 26 on its v5 index, behind DeepSeek V4.1 Flash and GLM-5.3 at 32, The Decoder reports. It beat Claude Opus 5.5 in only 22% of Surge's blind matchups.
- Big iron to self-host — SGLang's docs list eight Nvidia B300 GPUs for the standard version, or four GB300s for a more compressed variant. It generates 161 tokens per second, per Artificial Analysis.
Teams that need open weights for coding agents, security triage or data-residency rules should add Large 4 to their evals next to GLM-5.3 and DeepSeek V4.1 Flash, and compare cost per task rather than per-token price. Read the license if you resell hosting: commercial cloud services need a separate Mistral license. Kimi K3.1 Pro's weights, due by November 4, could reshuffle the open leaderboard within days.
🏢 DOJ asks if Hugging Face would give Nvidia a key AI distribution hub
The US Justice Department is asking whether Nvidia's 12.9-billion-dollar purchase of Hugging Face would let it capture a foundational point of distribution for AI, Reuters reported. Nvidia expects the deal, its biggest conventional acquisition, to close in the first half of 2027.
- Part of a wider probe — The questions run alongside a DOJ examination of Nvidia that began under Biden, covering its acquisitions, investments, chip sales and whether it pressured customers, including its SchedMD purchase.
- The Groq precedent — The DOJ has also reportedly examined whether Nvidia structured its 20-billion-dollar Groq licensing deal to avoid merger review. The Hugging Face purchase is big enough to require a formal premerger filing.
- Small revenue, big leverage — Hugging Face had annualized revenue of about 150 million dollars, Reuters previously reported. Jensen Huang has said developers won't need Nvidia hardware to build on or deploy through the platform.
An inquiry isn't a lawsuit, and Reuters says it may lead to no action, but regulators are treating the model hub as infrastructure rather than a community site. If your pipelines pull weights or datasets from Hugging Face, mirror the critical ones and confirm you could switch registries. If the deal clears, read any conditions on neutrality toward other chip and cloud vendors closely.
Become An AI Expert In Just 5 Minutes
If you’re a decision maker at your company, you need to be on the bleeding edge of, well, everything. But before you go signing up for seminars, conferences, lunch ‘n learns, and all that jazz, just know there’s a far better (and simpler) way: Subscribing to The Deep View.
This daily newsletter condenses everything you need to know about the latest and greatest AI developments into a 5-minute read. Squeeze it into your morning coffee break and before you know it, you’ll be an expert too.
Subscribe right here. It’s totally free, wildly informative, and trusted by 600,000+ readers at Google, Meta, Microsoft, and beyond.
🗞️ AI Bytes
🤖 Gemini 4 Carbon tops the Artificial Analysis index, but only defenders get it
Google's Gemini 4 Carbon scores 59 on the Artificial Analysis Intelligence Index v5, first overall, two weeks after Argon. It's going to vetted cyber defenders first, with no date for developers, at an introductory USD 2/USD 10 per million tokens.
🧰 Claude Code's background-agent status lines can eat 16% of your spend
Every ~30 seconds, Claude Code resends each background agent's full transcript to its model just to write a 3–5 word status line; one team measured that at 16.4% of spend. There's no setting to turn it off, but a community preload script stops the timer.
🔒 Open-weight GLM 5.3 built Firefox exploits for about USD 1,000 each
Epoch AI had GLM 5.3 turn Firefox fixes visible in public code into working exploit chains before the updates reached users, with no human input. Security fixes now need to ship faster and stay private until they do.
📈 ChatGPT and Codex will share one token-based usage budget from December
From December, Plus, Pro and Business plans draw ChatGPT and Codex usage from one allowance measured in tokens, with more advanced models and harder tasks using more. Heavy Codex users will share that budget with chat, so watch the usage page before limits bite.
📡 Release Radar
- SGLang v0.5.23 — Low-latency mode makes single-user DeepSeek-V4.1-Flash decode up to 44% faster on GB300; MoE first tokens arrive 30–58% sooner on H200.
- Nemotron 3.5 Transcribe — One open checkpoint transcribes 40 languages, translates 31 and labels up to eight speakers on recordings up to an hour long.
- Pydantic AI v2.65.0 — Adds ConversationStore with conflict-checked saves, WebSocket transport for OpenAI Responses, and fallback to the next model when one refuses.
- Cursor Agent SDK — The framework behind Bugbot: define event-driven agents in code, like PR reviewers, and deploy them to Cursor's infrastructure with one command.
- llm 0.37 — Simon Willison's CLI adds GPT-6.1 Sol and Terra, commands for OpenAI-compatible endpoints, and makes GPT-6 Luna the default; drops Python 3.10.
📖 Worth Your 10 Minutes
The AI cost of JavaScript frameworks
Lawson had Opus 5.5 build the same app five times in each of 11 frameworks plus vanilla JS, measuring cost, code size and Lighthouse scores. Hono+htmx and Astro were cheapest to build and Angular cost nearly twice as much: evidence, not vibes, on "agent-friendly" stacks. (Nolan Lawson)
🛠️ Top AI Tools This Week
Floorplan
Ask Claude Code or Codex for a floorplan and get a zoomable map of your app: parts, files, endpoints and cross-repo requests, plus a health check for keys in code and open Supabase tables.
unslop
A Claude Code and Codex skill: audits AI-looking UI (purple gradients, glows, emoji labels, uneven spacing), proposes a design direction, then fixes the screen pass by pass. Works with React, Vue, Svelte or HTML.