{"slug": "ai-news-november-01-2026-mistral-large-4-drops-to-0-70-m-backdoored-claude-hits", "title": "AI News — November 01, 2026: Mistral Large 4 Drops to $0.70/M, Backdoored Claude Extension Hits Cursor", "summary": "Mistral released the full weights of its 1-trillion-parameter Mistral Large 4 and cut API pricing roughly 50% to $0.70 per million input tokens and $2.10 per million output tokens, making the model competitive where its preview was not. Separately, Anthropic's Felix confirmed on Hacker News that versions 2.1.312 and 2.1.313 of the Claude Code VS Code extension were tampered with on Cursor's Open VSX-based marketplace to append an attacker SSH key labeled \"svc-telemetry\" to users' authorized_keys, live for about eleven hours, and that the incident was not a compromise of Anthropic. Cisco Talos also documented VOLTA-AUTODOXER, a Windows infostealer that harvests credentials and crypto wallets and POSTs the loot to OpenRouter, defaulting to Claude with a three-model fallback chain, to rank victims by value.", "body_md": "Good morning. Mistral’s trillion-parameter release is still rippling through the week, but today’s throughline is Anthropic playing both villain and victim — a backdoored Claude Code extension on Cursor’s marketplace, Claude’s safety classifiers blocking legitimate developer work, and new restrictions on game reverse-engineering that have HN commenters pointing out some obvious irony. Meanwhile, an infostealer is now using Claude via OpenRouter to appraise its victims.\n\n**Mistral Large 4’s weights are out.** As we covered yesterday, [Mistral released the full weights](https://mistral.ai/news/mistral-large-4-oss/) of its 1T-parameter “Le Chonk,” and the HN thread has now settled on a clearer verdict: the ~50% price cut to $0.70/$2.10 per million tokens makes it genuinely competitive where the preview wasn’t. The licensing fine print still trips up anyone running a model-as-a-service, and self-hosters with four H200s are reporting they’d need to drop to 3-bit or 4-bit quantization just to fit it.\n\n**A backdoored Claude Code extension hit Cursor’s marketplace.** [Phorion’s writeup](https://phorion.io/blog/offical-claude-code-vsx-compromise/) details how versions 2.1.312 and 2.1.313 of the Claude Code VS Code extension were tampered with to append an attacker SSH key labeled `svc-telemetry` to users’ `authorized_keys`, with persistence hooks across macOS, Linux, and Windows. Anthropic’s Felix clarified on HN that this wasn’t a compromise of Anthropic — the attacker uploaded a tampered copy to Cursor’s Open VSX-based marketplace, which was live for about eleven hours. A macOS bug limits inbound SSH impact for standard users, but anyone who installed those versions should still grep their SSH keys.\n\n**Anthropic is blocking users whose prompts touch model training.** A researcher claimed on Twitter he was [flagged for “training a competitor”](https://twitter.com/jiahaoshao1/status/2116475922190922088) despite only automating job submission with Claude. The broader HN thread catalogues escalating friction with Claude’s classifiers: refusals to refactor a user’s own monolith, blocks on firmware work for personally-owned devices, and injected moral commentary in generated artifacts. One commenter noted the affected user is in Hong Kong accessing via VPN, which likely tripped the flag — but the consensus takeaway was simpler: build provider-agnostic workflows now.\n\n**Anthropic also moved to block Claude from reverse-engineering games.** [Mashable reports](https://mashable.com/tech/anthropic-takes-steps-to-block-video-game-cloning-and-reverse-engineering) the restrictions follow viral incidents like an AI-assisted PS5 jailbreak and browser ports of Vice City. HN was not kind: “If they didn’t want us to use other people’s ideas without permission then why did they use other people’s ideas without permission to make the ideal tool for using other people’s ideas with permission?” Others pointed out the block appears to be a system-prompt injection easily stripped with a proxy, suggesting it’s optics more than enforcement.\n\n**Malware is now using Claude to appraise its victims.** Cisco Talos [published a breakdown](https://blog.talosintelligence.com/appraisal-by-ai-model-driven-victim-prioritization-in-the-volta-infostealer/) of VOLTA-AUTODOXER, a Windows infostealer that harvests credentials and crypto wallets, then POSTs the loot to OpenRouter — defaulting to Claude, with a three-model fallback chain tuned to dodge refusals — to generate a ranked dossier of what each victim is worth. The operational tell: the OpenRouter API key sits on the victim’s machine. Commenters flagged the obvious next step: proportional targeting, where high-value victims get zero-days and deepfakes while the rest get a scripted shakedown.\n\n**An AI agent injected drugs into a live mouse’s brain.** [Wired reports](https://www.wired.com/story/an-ai-agent-did-brain-surgery-on-live-mouse/) that Exin Therapeutics had Claude Sonnet 3.5 autonomously pick a treatment from 19 options, identify coordinates, and operate a robotic injector on a live mouse. The detail worth sitting with: Claude Opus 3.5 refused to complete the procedure mid-experiment, citing ethics. The more capable model was the one that balked.\n\n**ChatGPT account takeover via a shared link.** Parameter researchers [chained ChatGPT’s HTML Preview, a MathLive postMessage listener, and JSONP](https://www.parameter.ai/labs/we-found-a-gap-at-openai) to escape the sandboxed iframe and execute JS in a victim’s authenticated session — triggered by nothing more than opening a shared conversation link. OpenAI patched it October 26 and classified it P1, then paid out $5,000. HN’s reaction was unanimous that the bounty is insulting for a one-click full-account takeover, and worried others won’t bother reporting next time.\n\n**A five-model Qwen team beat Fable 5 on LiveCodeBench Hard.** [Memdoor](https://github.com/guregodevo/memdoor/tree/main/examples/workflows/team-solve) orchestrates five Qwen3.8-27B instances in specialized planner/ideator/solver/manager roles communicating via disk files, scoring 9.8/10 on the first ten Hard problems versus Fable 5’s 8.0, at $0.39 per run. Caveats are large — ten problems isn’t a benchmark, LiveCodeBench Hard is old enough to be saturated, and training-data contamination is unverified — but it adds to the case that harness design is the next scaling axis for anyone trying to stay off hyperscaler APIs.\n\n**GPT-5.6 Sol finished Factorio on peaceful mode.** Ryan Madden [documented](https://ryanmadden.net/gpt-56-beats-factorio/) a 13.5-hour run where Sol launched a rocket, though commenters noted it never really grasped belts — it manually ferried supplies between factory sections rather than connecting them. Peaceful mode removed the pressure of bug attacks, and the model had access to every Factorio guide ever written, so take the headline accordingly. One observer did note the factory designs look genuinely alien, which is at least evidence of synthesis over copying.\n\n**Leo de Moura on Lean and AI-assisted formal verification.** The Lean creator (now at AWS) [talks with The Data Exchange](https://thedataexchange.media/leo-de-moura-lean-aws/) about verifying AI-generated code and shifting developer effort from implementation toward specification. The sharpest HN question was also the obvious one: if an agent writes the spec and the proof, how do you know the properties being verified are the ones you actually wanted?\n\nThat’s the morning. If you’re running the Claude Code VS Code extension on Cursor, go check your `authorized_keys` before your second coffee.", "url": "https://wpnews.pro/news/ai-news-november-01-2026-mistral-large-4-drops-to-0-70-m-backdoored-claude-hits", "canonical_source": "https://ai0.news/posts/2026-11-01-daily-digest/", "published_at": "2026-11-01 07:00:07+00:00", "updated_at": "2026-11-01 07:20:42.886274+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-safety", "ai-agents", "ai-tools"], "entities": ["Mistral", "Mistral Large 4", "Anthropic", "Claude Code", "Cursor", "Felix", "Cisco Talos", "VOLTA-AUTODOXER"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/ai-news-november-01-2026-mistral-large-4-drops-to-0-70-m-backdoored-claude-hits", "markdown": "https://wpnews.pro/news/ai-news-november-01-2026-mistral-large-4-drops-to-0-70-m-backdoored-claude-hits.md", "text": "https://wpnews.pro/news/ai-news-november-01-2026-mistral-large-4-drops-to-0-70-m-backdoored-claude-hits.txt", "jsonld": "https://wpnews.pro/news/ai-news-november-01-2026-mistral-large-4-drops-to-0-70-m-backdoored-claude-hits.jsonld"}}