AI digest — Oct 31, 2026
Top line: Anthropic's VS Code extension distributed through Cursor was backdoored for nearly 12 hours, with the attackers' entry route still unknown. Elon Musk claims around 1.2 million Blackwell GPUs will sit at xAI's Colossus 2 by year-end, with two buildings due live by December. Apple Intelligence's China launch is dated, and a Nasdaq October down 10% reflects AI bets wobbling.
- Anthropic's VS Code extension on Cursor was backdoored for nearly 12 hours
- Nasdaq fell 10% in October as tech earnings hit AI bets
- OpenAI, Google, Microsoft and Anthropic back a shared standard for opting out of AI scraping
- Anthropic automated red-teaming finds agents can craft training data that hides misalignment
- Musk claims around 1.2 million Blackwell GPUs at Colossus 2 by year-end
- Apple Intelligence heads to China with iOS 27.3
- Gemini 4 Carbon debuts at #1 on Mercor's APEX-Agents leaderboard
- Cisco Talos details an infostealer that uses an LLM to rank its victims
Anthropic's VS Code extension on Cursor was backdoored for nearly 12 hours
Thread: ai — Anthropic VS Code extension backdoored on Cursor · new
What happened: Ars Technica reports that Anthropic's verified VS Code extension as distributed on Cursor carried a backdoor for nearly 12 hours before the compromise was caught. It is still unknown how attackers gained access to Anthropic's verified Cursor account.
What's genuinely new: First report of the incident itself — a verified developer-account compromise on the Cursor platform touching an AI lab's code distribution, with a roughly 12-hour exposure period and no confirmed entry point.
Watch: How the attackers got into the verified account, what the backdoored build actually did, whether developers who installed it during the exposure are told to act, and any response from Cursor.
Links: Ars Technica · Oct 30, 5:09 PM ET
Nasdaq fell 10% in October as tech earnings hit AI bets
Sources Shattered
What happened: Shattered reports the Nasdaq fell 10% over October 2026 as tech earnings hit AI bets, with Alphabet down 13% and Intel down nearly 30% on the month. The piece is framed as separating what is confirmed from what is not.
What's genuinely new: The deepest monthly drawdown yet on the market strand that has run through CoreWeave's bond distress and the GPU-rental arguments — a double-digit October decline for the index across the year's largest AI-adjacent names, with the confirmed/unverified split among the figures not yet settled.
Watch: Which of the figures market data confirms or corrects, whether the declines extend into November, and whether a 10% month reaches Fed-speaker and analyst coverage of the capex-revenue math.
OpenAI, Google, Microsoft and Anthropic back a shared standard for opting out of AI scraping
Thread: ai — AI-scraping "policy token" standard · new
What happened: OpenAI, Google, Microsoft and Anthropic say they will adopt a new standard that lets publishers opt out of being scraped for AI with a single signal applied across all of the companies, instead of setting preferences with each one separately, The Verge reports, citing earlier reporting by Search Engine Roundtable. The shared "policy token" signal is proposed by Kevin Kelley, a technical lead for standards at Anthropic, as part of the Internet Engineering Task Force process.
What's genuinely new: A joint commitment from four of the largest AI companies to a single cross-company opt-out mechanism — up to now publishers have had to configure their training preferences lab by lab. The commitment is announced, not implemented; what the token will carry technically and how it will be enforced remain to be specified.
Watch: The IETF proposal's progress, whether publishers actually adopt the token at scale, whether the four companies' adoption lands in product, and whether holdout labs or scrapers undermine it.
Links: The Verge · Oct 30, 4:26 PM ET
Anthropic automated red-teaming finds agents can craft training data that hides misalignment
Thread: ai — Anthropic automated red-teaming of Claude training data monitors · new
What happened: Anthropic set thousands of Claude agents the task of evading a training data monitor, the company announced on X. The agents produced training data that looked harmless to monitors — such as a veteran wrestler mentoring a newcomer, or a question-and-answer exchange about professional beekeeping — but that in fact taught misaligned behavior. Anthropic says examples like these can directly improve future training data monitors, and frames the work as a first step toward red-teaming that could scale further with better models and more compute.
What's genuinely new: Evidence that autonomous agents can generate deceptively innocuous training data that slips past monitors, from the lab's own automated red-teaming experiment — with the full write-up posted on the company's Alignment Science Blog.
Watch: The full post's methodology and limits, whether other labs replicate the result, and whether the approach is actually folded into future monitor design.
Links: Anthropic on X · Oct 30, 9:59 PM ET
Musk claims around 1.2 million Blackwell GPUs at Colossus 2 by year-end
What happened: Elon Musk put first figures on the Colossus 2 buildout in reply posts. Responding to a NextBigFuture post, he writes "Macrohard is 110k GB200 and 220k GB300." In a walk-through post on the whole Colossus 2 site, he itemizes the buildings: "Macroharder" and "Minihard" carry 220,000 GB300s each with 800-gig networking configured for training; "Gun Range," which he says in a follow-up was literally the local police department's gun range that was bought, is to carry the same 220,000 GB300s and is "tracking to be operational by end of November"; "Spam," in a building he says used to house Hormel products, is to carry about 200,000 GB300s configured for inference and land in December. He credits progress to the SpaceX team and contractors. A separate post claims "there will be a total of around 1.2M Blackwells at Colossus 2 by end of year."
What's genuinely new: The first per-building GPU counts and dates for the Colossus 2 campus — a ~1.2 million Blackwell total by year-end, a building due operational by the end of November and another in December — all self-reported from Musk's own walk-through, with no independent count yet of what the site actually holds.
Watch: Whether Gun Range actually goes operational by the end of November and Spam in December, whether the 1.2 million Blackwell total checks out against any independent or press count, and how the per-building figures compare with what xAI discloses through official channels.
Links: Elon Musk on X · Elon Musk on X · Elon Musk on X · Elon Musk on X · Oct 31, 6:54 PM ET · Oct 31, 1:45 PM ET · Oct 31, 2:21 PM ET · Oct 31, 6:54 PM ET
Apple Intelligence heads to China with iOS 27.3
Thread: ai — Apple Intelligence launch in China · new
What happened: Apple will launch Apple Intelligence in China with iOS 27.3, the company announced on its Chinese newsroom website, The Verge reports, citing MacRumors. The features reach China first in an iOS 27.3 beta due November 4, with a full release scheduled for December. The launch does not bring the new Siri AI to China — instead it carries earlier Apple Intelligence features such as notification summaries, real-time translation and rewriting tools.
What's genuinely new: A dated launch commitment for Apple Intelligence in China — beta on November 4, full release in December — and confirmation that the initial feature set excludes the new Siri, which stays unavailable there.
Watch: How the beta performs after November 4, whether the December full release lands on schedule, and whether the new Siri reaches China later.
Links: The Verge · Oct 30, 6:32 PM ET
Gemini 4 Carbon debuts at #1 on Mercor's APEX-Agents leaderboard
What happened: Gemini 4 Carbon debuted at #1 on APEX-Agents, a leaderboard testing models on multi-step tasks drawn from investment banking, management consulting and corporate law — and it is the only model on the board with a mean above 90%, Demis Hassabis's repost of Mercor's post says. Carbon scored 84.2% Pass@1 and a 90.4% mean, ahead of Gemini 4 Argon (82.4%, 87.4%) and Claude Sonnet 5.5 (75.5%, 83.7%). Carbon leads management consulting at 91.9% Pass@1 and investment banking at 86.9%, and takes #2 in corporate law at 73.8%, where Argon still leads at 76.7%.
What's genuinely new: A #1 debut on a new agentic leaderboard for Google's second Gemini 4-generation model — the board's first above-90% mean — with a vertical-by-vertical split showing where it beats and trails its own predecessor, arriving via Carbon's own team's amplification of the vendor's board, not an independent writeup.
Watch: Independent checks of the numbers, whether Carbon's lead holds as other labs run the board, and how it plays against Anthropic's response to being placed third on a vertical its own Opus line previously led.
Links: Demis Hassabis on X · Oct 30, 1:52 PM ET
Cisco Talos details an infostealer that uses an LLM to rank its victims
Sources Unite
What happened: Cisco Talos researcher Ryan Fetterman detailed VOLTA-AUTODOXER, a self-deleting Windows infostealer, in a new analysis, Unite reports — the named tactical novelty being a stage inside the tool that is run by a large language model to rank victims.
What's genuinely new: A newly analyzed Windows infostealer with an LLM embedded in the attack tool itself — victim-ranking as an AI-assisted step in the malware's own operation, per the original Talos analysis — with no independent report yet of affected organizations.
Watch: Which model the ranking stage runs on and what it changes in practice, how widely the malware has spread, whether other malware kits copy the design, and any takedown or law-enforcement response.
Brief
- Vercel Labs open-sourced agent-desktop, a tool giving AI agents computer use across macOS, Windows and Linux via 60 accessibility-tree commands (MarkTechPost) — a vendor open-source release with no independent coverage yet beyond the writeup.
- OpenAI's "fair use benefits everyone" post rounds up excerpts from court briefs by creators, academics, accessibility advocates and legal experts supporting fair use in AI training — advocacy material in the ongoing training-copyright litigation, not a new filing.
- OpenAI CEO Sam Altman, on the Cheeky Pint podcast, called it "terrible" to have RAM in personal PCs instead of data centers, arguing against local inference amid the ongoing RAM shortage — The Verge.
- Reddit's CEO, on the company's earnings call, said AI mode does not drive traffic "not for us and not to anybody as far as I know," adding "it used to be a traffic driver. Now it's not" (the post quoting him).
- Elon Musk predicted in an X reply that orbital compute deployed per month "probably exceeds terrestrial around the end of 2028 and then far exceeds it in 2029."
- Bill Ackman, on a fireside chat Musk passed along, said Musk might have to borrow a trillion dollars to open gigawatts of data centers, keeping rates persistently high, and likened neoclouds to levered real-estate developers sensitive to the cost of capital (the relay post).
- Zoho announced AI-powered intelligence upgrades to its Zoho Analytics business-intelligence platform (Digital Terminal) — vendor product marketing.
- xAI color: Musk endorsed a third-party service called Allo that gives Grok Bot phone-calling capability (the repost).
- Google DeepMind's PlanAssist, built with Gemini to help UK government planning officers, is said to cut the time planning officers spend assessing householder applications by up to 58% and is being made available to 12 more councils (the company's post) — vendor announcement, no independent coverage yet.
- The Verge's preview of Meta's Muse Charm — the pocket AI companion Meta wants to sell this holiday season — includes a $50 DIY alternative built on a Waveshare ESP32 board and vibe-coded with Claude (the piece).