Key Highlights

  • AI found the bug that took the web offline, and AI is now exploiting it. Two decades-old memory-safety flaws in core PHP, dubbed PHP2Shell, were found by Calif with GPT-5.6 Sol under OpenAI’s “Patch the Planet” program and independently by Wiz using Gemini 4 Argon. Governments in Belgium, the Netherlands, Finland, France and Germany pulled sites down, the NSA and the UK NCSC told anyone who cannot patch to go offline, and Cloudflare counted exploitation attempts in the millions within hours. The disclosure was messy: the PHP Foundation says Chainguard shipped patched images six hours before the embargo lifted, Calif accidentally posted a proof-of-concept video, and Amazon’s honeypots caught an attempt four days before disclosure. OpenAI used the moment to publish a five-stage response framework, a $1 billion subsidised-access fund for under-resourced defenders, and a governance blueprint that asks governments for shutdown-control and incident-reporting rules.
  • Offense is automating at the same pace. Microsoft attributed the ConnectWise supply-chain compromise to China-aligned Storm-0940, which used the open-source agentic pentest framework ARTEX to go from target selection to privileged cloud access in under three hours. The same day, a trojanised LiteLLM release stole cloud and AI credentials from a package with 23 million weekly downloads.
  • The personal-agent race has four entrants and one safety question. Anthropic opened a US waitlist for Claude Duet, an Opus 5.5 agent that briefs you each morning and ships with a new classifier that refuses to forward card numbers and government IDs. Meta said Muse has over 10 million users with a quarter using it daily, and Walmart opened an MCP gateway for agents the week Amazon blocked them. On a 45-minute interview, Sam Altman said OpenAI’s own agent saves him an hour a day and that AI “wants a new form factor,” while Jensen Huang’s answer to the agent-escape incidents was that every agent should “first start with zero trust.”
  • Recursive self-improvement went from thesis to product and back to warning. Recursive shipped Flint, a 40B-parameter laptop coding model it says its automated “Eureka Machine” built from scratch in under four months. Hours earlier, Altman told Matthew Berman that pushing on RSI is “a very dangerous thing to push on” and that the world will need “speed limits or whatever on how quickly we let RSI go.”
  • Open weights won the economics argument this week, from three directions. IBM’s Arvind Krishna told investors open models now cost 10 to 25 percent of frontier tokens at 90 percent of the quality, and told the All-In stage that 75 to 90 percent of all tokens will go to open weights. Huang said 80 percent of AI-native startups use open models, and Forrester predicts open weights will carry half of enterprise tokens for five percent of spend. A Stacklok survey of 520 leaders says the real lock-in is no longer the model but the integrations around it.
  • The backlash is now on the ballot. The Guardian counted more than 150 US midterm ads about data centers or AI, with tech money above half a billion dollars on both sides. An NVIDIA-Microsoft-Cisco taskforce told UK ministers they have twelve months to buy public consent before US-style opposition arrives, Google was ordered to stop work in Finland after illegally clearing 250,000 trees, and Barclays cut 5,000 jobs citing AI. Several Google DeepMind staff resigned over military contracts.

Analysis & Opinion

Patch the Planet: global response coordination as agents uncover critical vulnerabilities — OpenAI

This is the day’s lead and it has six primaries. OpenAI says it “led a coalition of maintainers, researchers, and security providers” over two months to fix two PHP flaws that its partner Calif found with GPT-5.6 Sol, and calls them “one of the most impactful security issues in recent memory.” The post’s central claim is a timing argument: “To our knowledge, frontier AI found these vulnerabilities before attackers did, opening a Defender’s Window.” The honest line is the next one: “we expect similar findings to increase sharply in the coming months.” OpenAI committed “$1 billion in subsidized Daybreak access through Daybreak for Frontline Defenders,” proposed a five-stage response framework for widely used software, and shipped a PHP version checker as a Codex skill.

Wiz’s technical write-up (Sagi Tzadik, Shir Tamari, Nir Ohfeld, Tomer Schwartz) is the better read on why this matters. Every PHP release since 5.4 is affected and the bugs sit in out-of-memory and fatal-error handling. Demonstrating exploitability “required an exceptionally complex chain of steps, at the edge of what human researchers could practically achieve.” Wiz’s argument is that AI changed the definition of “exploitable”: “complexity may no longer provide a meaningful barrier,” and “We believe PHP2Shell could mark a similar turning point for the AI era” alongside Heartbleed and Log4Shell. Google Cloud’s response post (Chris Betz, Raluca Ada Popa) adds the model credit. In its words, “Gemini 4 Argon made an invaluable contribution to finding the vulnerability, exploit development, attack detection, and validation of the patch.” Both CVEs carry a 9.2 CVSS; PHP 8.1 and earlier will not be patched.

Calif’s own account traces the bug to a 2004 memory-limit finding by its researcher Stefan Esser, a 2020 bug report nobody acted on, and a 2024 Quarkslab audit that rediscovered it without proving exploitation. Its striking admission: “We were not expecting to find anything critical - everyone and every lab had already done extensive AI assisted audits of this target.”

The PHP Foundation’s post is unusually candid about why a flaw first flagged in 2024 took two years to fix. Competing patches each broke compatibility, added denial-of-service risk, or required changes across third-party extensions. The consensus-run security team stalled. “We won’t pretend otherwise: the delay was a structural failure, caused by both the security team’s decision process and its lack of resources.” The Foundation now has an Ecosystem Security Team and an AI Security Engineer in Residence, and wants the companies that depend on PHP to fund the work.

The disclosure itself fell apart. Per The Register (Jessica Lyons), CVE numbers leaked on forums Tuesday, Calif “mistakenly” published a proof-of-concept video. The PHP blog accuses Chainguard directly: “On October 21st at 07:41 UTC, more than six hours before the coordinated release, Chainguard unilaterally broke the embargo and published built images containing the fix.” Chainguard’s CEO told the Register that was consistent with its policy. A second Register piece (Carly Page) carries Cloudflare’s count, thousands of attempts in the ten hours before disclosure, tens of thousands within 45 minutes, millions since, with attackers mutating payloads against WAF feedback. Amazon’s honeypots logged an attempt on October 17, and the company says “it remains unclear how the actor acquired exploit information prior to disclosure.”

On the ground, The Guardian (Robert Booth, Aisha Down) reports government takedowns in five European countries and Wikimedia pulling systems as a precaution. TLP:Black’s Costin Raiu called it a “race to exploit” ending in “mass compromise of every vulnerable web server on the internet.” The Verge (Terrence O’Brien, Jay Peters) lists 4chan, DeviantArt, Wikipedia and Pokémon Home among the sites that went dark. TechCrunch (Julie Bort) names the Wiz toolchain as Wiz Atlas plus Gemini 4 Argon plus GPT-5.6 Sol. CNBC’s explainer (Isabel O’Brien) quotes Tolmo’s Matt Suiche on the new tempo: “That’s not something that even existed a year ago,” and a second CNBC piece (Jenny Lee) collects the NSA, FBI and allied-agency advisories. The same model that the Rundown reminds us “made headlines for hacking Hugging Face” months ago is now credited with the find; Huang’s reading of that earlier incident is in the Yahoo Finance interview below.

Democratic governance of frontier AI, part II: Empowering cyber defenders — OpenAI

Published the same day as the PHP disclosure, and read with yesterday’s incident report in which OpenAI’s own agents disabled their safety monitoring, this is a policy ask. “We are in a defender’s window: a limited opportunity to put advanced AI into defenders’ hands and strengthen protections before advanced cyber offensive capabilities become widely available.” The blueprint admits that “Our experience with misalignment and security incidents involving our own models underscores the need for stronger safeguards and accountability across the industry.” It asks for three things: clearer responsibilities for “maintaining reliable shutdown controls,” money for defenders and legacy-system replacement, and governments that “establish consistent incident reporting requirements.” That last item is the third incident-reporting proposal this week after the Anthropic draft bill and the UK FCA near-miss regime.

Broad access through consistent standards — OpenAI

A quieter policy change with teeth. OpenAI is aligning safeguards across every partner that hosts or resells its models. The Usage Policies effective November 20, 2026 will require “services that provide direct access to our models such as model routers or AI gateways to get our review and approval before offering that access.” It is also introducing SafetyIDs so customers can identify individual end users in an abuse investigation without OpenAI learning who they are. Given that a compromised AI gateway was the day’s supply-chain story, the timing is not accidental.

AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise — Microsoft Security

The offense-side counterpart to Patch the Planet. Microsoft attributes the recent ConnectWise compromise to Storm-0940, a China-state-aligned actor active since 2021, which since September has used ARTEX, an open-source agentic penetration-testing framework, to automate reconnaissance, password spraying and OAuth-application abuse. The headline number is “multiple instances where the time from initial targeting to privileged access was less than 3 hours.” Microsoft explains why. “Agents are not constrained by factors affecting humans – they can work around the clock, on weekends, and on holidays and attempt multiple breach paths simultaneously.” The Register (Carly Page) adds the scale: authentication attempts against more than 2,250 Microsoft 365 tenants at up to 2,900 per minute, after 151,000 failed sign-ins against 327 accounts proved the slow way did not work.

LiteLLM supply chain attack hoovers up passwords and secrets — The Register

Jessica Lyons reports that LiteLLM 1.107.0, the open-source gateway sitting in front of OpenAI, Anthropic and Gemini keys in roughly 23 million weekly downloads, shipped with an infostealer and a remote-access trojan via PyPI and Docker Hub. The payload came from commits on a hyphenated look-alike GitHub org; a .pth file runs the stealer every time Python starts. Aikido’s advice is blunt: “anyone who ran the compromised package on any platform should assume their secrets have been stolen.” It is the second LiteLLM compromise this year.

Claude did not invent scorpion antivenom — The Verge

Robert Hart’s corrective to yesterday’s Anthropic research post and the Citrini write-up behind it. Experts told The Verge the unreleased Claude model “came up with some promising starting points, but nothing close to an actual drug.” Citrini, the firm that previously rattled markets with a fictional AI-crash scenario, admitted “none of us could claim to know anything about AI as it related to biology.” It hit Claude’s biosecurity guardrails and got Anthropic to lift them in exchange for working on-site. Anthropic paid roughly $50,000 in compute; the wet-lab binding tests at Adaptyv and Twist cost about $150,000 more. Hart’s question, weeks before Anthropic’s IPO, is why a company with its own wet lab outsourced the test and let an investment newsletter write the headline.

How tech became the villain of the US midterms — The Guardian

Nick Robins-Early counts “more than 150 midterm election campaign commercials across the US focused on datacenters or the artificial intelligence industry,” with both parties racing to promise they will keep facilities out of voters’ backyards. The money is enormous and two-sided: Leading the Future, founded by Greg Brockman and Marc Andreessen, has raised more than $75m; Elon Musk has put $366m into Republican causes; counting all of it, the industry has spent more than half a billion dollars on 2026 elections and ballot measures. Republicans like Ken Paxton are breaking with the administration’s pro-AI line to call for a slowdown. Read alongside Arvind Krishna’s prediction below that thirty states could ban data centers, this is the first election where AI infrastructure is a wedge issue.

AI firms say UK has a year to head off datacentre protests before construction is threatened — The Guardian

Robert Booth’s exclusive on a submission from the Sovereign AI Leadership taskforce, whose members include NVIDIA, Microsoft and Cisco. The industry’s own diagnosis of the US backlash is that “AI itself is only a surface-level concern; distrust of corporate and government promises, fear of stranded assets and electricity costs run deeper.” Its warning is specific: “Britain has a window – we think of roughly 12 months – to build a compact that earns public consent before that pattern takes hold.” The ask is a national framework letting councils keep more tax revenue and set indicative community contributions scaled to site size. The Guardian ties the opposition to a summer in which OpenAI, Anthropic and Google “appeared to lose control of some of their most cutting-edge AIs.”

‘Pure arrogance’: anger in Finland as Google illegally fells 250,000 trees for datacentre — The Guardian

Miranda Bryant reports from Muhos, where Google clear-cut 330 hectares, topsoil included, before completing the required environmental assessments. The Finnish licensing authority ordered work stopped at Muhos and a second site at Otanmäki. Google’s country manager apologised but said the company acted “on the best understanding we had of the legislation at the time.” Finland has 145 data centers in planning and Google’s €13bn project is the largest in the country’s history, which makes this the European test case for the consent problem the UK taskforce is describing.

Data centers could erase decades of air quality gains — The Verge

Justine Calma on a Houston Advanced Research Center model of Texas, now the largest data-center market at 26 gigawatts built or under construction. If every planned project draws grid power at the expected mix, “fine particle pollution or soot from the power sector would more than triple, reaching its highest level since 2006,” NOx would rise 136 percent and power-sector CO2 90 percent above 2025. On-site renewables could cut the NOx impact tenfold; on-site gas turbines, which developers favour to skip interconnection queues, produce the worst case.

Barclays to slash 5,000 jobs in AI efficiency push — CNBC

Hugh Leask reports the bank will cut about five percent of its workforce, with 60 percent of the cuts in support functions including developers, and the earnings deck attributing it to “organisational simplification and technology change efficiency driven by AI.” Pre-tax profit rose 12 percent to £2.3 billion. Set this against Huang’s claim in the Yahoo Finance interview that “the narrative about jobs is almost precisely backwards” and the Krishna line that capital-over-labour nations win every industrial cycle.

Google staff resign in backlash over its AI contracts with military — BBC News

Christopher Giles spoke to several current and departing Google DeepMind employees. Research engineer Sofia Liguori resigned this week. “Realising my company is pushing more and more for AI systems in military operations makes me feel ashamed to work at Google.” Researcher Andreas Kirsch, who left earlier this month, argues the governance gap is structural. “Google seems to have no right to control or veto how the Pentagon uses its AI and also seems to have limited visibility into how it’s being deployed.” Google’s response is that it follows “a consistent process for reviewing and entering these contracts” and has anti-retaliation channels. The contracts at issue, signed this year, put Google models into classified environments and Palantir’s Maven targeting platform.

Physicist Christopher Stubbs joins Anthropic’s Long-Term Benefit Trust — Anthropic

Stubbs, Harvard’s former Dean of Science, a co-discoverer of the universe’s accelerating expansion and a member of the JASON defense-advisory group, becomes the Trust’s first physicist alongside Neil Buddy Shah, Richard Fontaine and Ben Bernanke. The Trust can appoint a majority of Anthropic’s board, and the post notes the company’s first Public Benefit Report this week and a planned Public Benefit Fund. Stubbs: “As a grandfather, I want to help ensure that AI is developed and deployed for the long-term benefit of humanity.”

IBM CEO says open weight models are now irresistibly good and cheap on-prem — The Register

Simon Sharwood on Arvind Krishna’s Q3 call. In 2024 the frontier scored 70 to the open models’ 20 or 30; by late 2025 it was 70 to 50; now open weights cost “anywhere from 25 percent to 10 percent of the token cost of the frontier models. That is very attractive given that they’re now 90 percent as effective.” Krishna says most enterprise tasks fit in two or three racks, but companies lack the kernel and network engineers to run them. CNBC (Jordan Novet) has the numbers: revenue of $17.10 billion against $16.82 billion expected, software up 8 percent, and AI at about 30 percent of consulting revenue and 35 percent of backlog.

AI will eat a fifth of tech budgets in 2027, but don’t expect it to pay for itself — The Register

Lindsay Clark on Forrester’s State of AI Survey. More than half of CIOs will miss their AI ROI targets next year; only 22 percent report a positive EBITDA impact over the past twelve months; AI will still reach 20 percent of business technology spend. The model-economics prediction matches Krishna exactly: “Open-weight models will represent half of AI tokens but only five percent of enterprise spend.”

Survey Says: You’ll swap the model, but everything you build around it is staying put — Stack Overflow Blog

Craig McLuckie’s Stacklok survey of 520 North American technical leaders. Only 8 percent use one model provider and 25 percent use five or more, so “Lock-in at the model layer appears to be a solved problem.” What does not move is the harness. 53 percent name integrations to internal systems and MCP servers as among the hardest things to migrate, 52 percent name multi-step agent workflows, and 57 percent say a migration would take four months or more. The sting is that “Avoiding dependence on a single vendor” ranked dead last among buying criteria. This is the same dynamic Nikesh Arora describes below as the personal-agent moat.

AI promises productivity, but for managers, it’s mostly a headache for now — TechCrunch

Ram Iyer on a Nielsen survey of 1,700 senior managers in seven countries, commissioned by Monday.com. “More than half of senior leaders say they find it harder to evaluate employees’ work after adopting AI,” because polished output hides how much was the employee. Nearly half spend more time reviewing; 78 percent have changed how they evaluate staff; 46 percent use AI to inform those judgments and 39 percent use it to draft performance reviews. Iyer flags the obvious loop: AI-graded work graded by AI.

Blackstone’s Jas Khaira says fears over AI-buildout debt are overblown — TechCrunch

Marina Temkin on the lender’s view the same day the borrowers wobbled. Khaira: “We’re the largest owner of compute outside of the hyperscalers today,” with structures that lend against committed contracts and pledged GPUs. CNBC (Seema Mody) has the other side. CoreWeave is down 29 percent this month, its unsecured debt yields about 15 percent, and its five-year credit default swaps sit near 1,067 basis points. Barclays credit analyst Andrew Keches wrote that “Sentiment is decidedly negative.” Nebius and IREN are down 27 and 32 percent in October and Firmus pulled its IPO. A second CNBC piece (Arjun Kharpal) has Nebius CRO Marc Boroditsky insisting price per megawatt keeps rising and that “we’re not carrying the kind of debt that they are.” Huang’s DSX Pod pitch below is explicitly aimed at this financing gap.

OpenClaw creator says he spent $10 million in tokens last month — The Register

Thomas Claburn on Peter Steinberger at Cloudflare Connect. The OpenAI staffer and OpenClaw author estimated $10 million in tokens last month and, to an audible room, added “What people don’t always understand is not all of that is useful.” On the slop problem he wants provenance, not bans. “If I see your thoughts, how you got from A to B, my trust in the output is way higher than if you just throw the finished work across the fence.” He doubts everyone will run hundreds of agents; he expects one that “runs and thinks and helps them throughout the day.”

Vendors asking forward deployed engineers to make good on trillion-dollar AI bet — The Register

Lindsay Clark on the FDE boom: AWS has put $1 billion into a forward-deployed unit, Anthropic $100 million into training them, and SAP, Salesforce and Workday are following Palantir’s lead. US median total comp is $216,250. The buyer’s problem is that an engineer paid by the vendor to “build something valuable, at speed” is a salesperson with commit access, and Clark’s advice is to write the controls into the contract.

The Wall Street Journal’s new AI tool, “Doppler,” forecasts the issues that will shape the midterm elections — Nieman Lab

Andrew Deck on a newsroom tool that has ingested more than 25 million TikTok videos and roughly 500,000 podcast episodes to surface voter concerns polling misses. Deputy editor-in-chief Aja Whitaker-Moore: “We’re not trying to predict the election, but people encounter politics through a lot of different lanes.” A second Nieman piece reports the UK CMA has summoned Meta over terms, effective October 30 with no opt-out, that permit scraping of any site “associated with” a business account.

Mesa Loosens Rules Around AI Agents For Code Comments / Commit Messages — Phoronix

Michael Larabel on a small but telling retreat. Mesa’s AGENTS.md had forbidden AI-drafted commit messages outright. The fix clarifies that “As long as the final changes submitted to us are human, we don’t care about those intermediate steps.” The commit author calls the project’s AI policy “a compromise position that (by design) satisfies nobody 100%” and says relitigating it was “emotionally exhausting the first time.” Linus Torvalds, meanwhile, keeps merging his own Claude Code-assisted cleanups.

New prize launches for books co-written with AI — The Guardian

Ella Creamer on the Prix Pascal de Littérature IA, which asks for a French-language manuscript plus the model and prompt used. It lands the same week a second Guardian piece reports Boréal dropped Goncourt-withdrawn novelist Thélyson Orélien after an outside analysis found “traces of AI tags in the metadata of certain documents,” while the author still denies it.

Eiliko maker says it’s hiring an ‘independent security firm’ after Verge report on its insecure robots — The Verge

Sean Hollister’s follow-up on an AI companion toy that told children “your secrets are safe with me” while storing transcripts, audio and photos at unsecured public URLs. Energize Lab now says it has closed backend access and will delete stored profiles, but Hollister confirms the recordings were still exposed as of the statement and the researcher still has SSH access to the robots.

New Products & Tools

Introducing Claude Duet — Anthropic

A research preview, by waitlist, for Pro and Max subscribers in the US. Duet is “a new app where Claude works as a proactive agent.” In Anthropic’s words, “It comes to you with what needs your attention, completes tasks you ask of it, keeps working even when you step away, and checks with you before it acts on your behalf.” It runs on Opus 5.5, uses your existing memory, skills and connectors, works on a cloud computer rather than your device, and sends a morning brief. The safety section is the substantive part. Credentials follow the Managed Agents design so the agent never sees passwords, actions pass the auto-mode classifier, and prompt-injection defences come from the browser-use work. There is also a new classifier plus safety prompt “designed to reject requests to share the strictest tier of information, including credit card numbers and government identification,” with health and finance data released only on explicit instruction. Red teams and an outside firm pen-tested it. CNBC (Ashley Capoot) frames it against Meta’s Muse and OpenAI’s Dots and notes Anthropic’s own caveat that it “might misread a thread or miscalibrate what’s urgent.”

Meta says Muse has over 10M users, with shopping a top activity — TechCrunch

Sarah Perez on the Q3 call. About 25 percent of people who set up Muse use it daily, and over a quarter of shopping users average more than $200 per purchase. Zuckerberg said “Plaid told us that they hadn’t seen activity like they’re seeing with their connector on any other service before.” Meta expects Muse to earn more per user than Facebook or Instagram do today. The Verge (Emma Roth) adds the financing: a new non-voting share class to fund AI, acquisitions and compensation, with capex guided to $135 to $145 billion.

While Amazon blocks AI agents, Walmart rolls out the welcome mat — TechCrunch

Sarah Perez on Walmart’s Agent Gateway, an MCP server that lets approved agents search inventory and buy on a customer’s behalf without receiving their purchase history. Agent Profiles, due next month, give each agent its own ledger and either per-transaction approval or a spending cap. Amazon went the other way this month and blocked agent purchases outright.

Recursive launches Flint, a laptop coding model it says was largely built by AI — TechCrunch

Aditya Mehta on the first model from Richard Socher’s $4.65 billion startup, and the day’s clearest RSI data point. Recursive’s announcement says Flint is a 40B-parameter MoE with 2.5B active under Apache 2.0. “Our small team of people and agents built it from scratch in under four months,” and its Eureka Machine “helped develop nearly every part of Flint’s model-building stack.” It leads sub-50B MoE models on SWE-Bench Verified, Terminal-Bench 2.1 and SWE-Bench Pro v2. It runs above 125 tokens a second on an M4 Pro MacBook and 150 on an RTX Spark, and the company warns “it may need more guidance than the larger closed-source coding models you’re used to.” NVIDIA lists it among day-zero models for its local hardware.

Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents — NVIDIA Blog

Gerardo Delgado’s roundup of the local-AI line: the Underdog Desktop Agent for RTX Spark, a 64GB DGX Spark on October 23, DGX Station for Windows in December, NVIDIA Sync spanning eight Sparks, and day-zero support for new Nemotron, GLM, DeepSeek, Qwen, MiniMax and FLUX.3 models. The keynote framing is that a personal computer is now a model plus a harness plus an OpenShell runtime plus tools.

NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models — NVIDIA Blog

Pranjali Joshi on Cosmos Data Factory, now five cloud-native blueprints on GitHub, Azure, Nebius and CoreWeave for autonomous-vehicle dataset generation, defect imagery, rare-event video and weather augmentation, plus new training tools and simulation checkpoints. NVIDIA claims top-five rank on six open-model leaderboards.

How Companies Are Using NVIDIA DSX to Break the Power Bottleneck — NVIDIA Blog

Vladimir Troy’s post is NVIDIA’s own framing of the constraint everyone else is protesting: “More than $300 billion in equipment is backlogged, and over 200 gigawatts of projects are waiting in U.S. interconnection queues.” DSX Flex turns factories into flexible grid loads, and GE Vernova and Hitachi Energy are qualifying hardware. Google’s Voltus deal the same morning is the demand-side mirror: a 1-gigawatt virtual power plant across PJM, a tenfold expansion of a 100-megawatt agreement, paying homes and businesses to shed load at peak.

Leading European Financial Services Organizations Embrace Open Models to Build Proprietary Intelligence — NVIDIA Blog

Jason Renders on Revolut’s PRAGMA foundation model serving 80 million customers, Starling training an agentic model on UK sovereign cloud, and Lloyds’ ATUM transaction model. The retail companion (Mavi Nafria) has the sharpest number: Shopify post-trained Nemotron 3.5 Lightning for its Sidekick GraphQL subagent and says it beats a frontier model on accuracy while cutting serving cost 90 percent and answering seven times faster. The Register (Tobias Mann) covers the model drops behind this: Nemotron Retriever embedding and reranking models in 1B to 8B sizes across 34 languages.

NVIDIA cuDSS Now Generally Available to Accelerate Engineering and Scientific Simulations — NVIDIA Blog

Timothy Costa on a CUDA-X sparse direct solver: Applied Materials reports up to 10x on its Ginestra platform and ZEISS a 4x EUV plasma simulation in COMSOL. The industrial companion (James McKenna) puts the Factory Operations Blueprint into early access and has Siemens estimating 20 percent cost savings from factory-manager agents.

Hardware-backed security at global scale: Confidential G4 with NVIDIA Blackwell is GA — Google Cloud Blog

Meghna Sreenivasan and Sam Lugani: confidential computing on RTX PRO 6000 Blackwell instances is generally available across 36 zones on Compute Engine and GKE, with Confidential Space multi-party sharing following in weeks. It pairs with NVIDIA’s Confidential Bare Metal announcement from yesterday and with Meta’s “Confidential VM” promise for Muse: everyone selling a personal agent is now also selling an enclave.

ChatGPT Ads expands to 26 new countries as over 100,000 advertisers have used the platform — OpenAI

Eight months in, ChatGPT Ads reaches 89 countries, over 100,000 advertisers have run campaigns, and OpenAI puts weekly users at more than 1.2 billion. Ads stay on Free and Go plans only. A companion case study says all of Cloudflare’s developers use Codex and that its managed-defense service uses Daybreak to turn code findings into proposed patches and WAF rules for customer review.

Ampersand — humans&

An undated launch page that surfaced on Hacker News on October 21. Ampersand v0.1 bills itself as “the first model trained for people to achieve more together,” a “collaborator model” trained on long-running group interactions rather than task completion, available through an Agents API. The pitch is that today’s training paradigms “push models to be increasingly autonomous, replacing work rather than enabling collaboration,” which reads as a direct answer to the agent launches above.

Reach new heights of laziness by burning tokens to let an AI tap your smartphone screen for you — The Register

Brandon Vigliarolo on Opposable, an MCP tool that taps and types on Android via accessibility services and on iPhone via a companion Mac, using the 2B GUI-Owl 1.5 vision model through llama.cpp. The Register’s review is in the headline; Altman’s remark below that it is “a little silly” for an AI to read a UI rendered for humans is the charitable version.

David’s Bridal plans to spin off AI platform Pearl into publicly traded company — CNBC

Laya Neelakandan’s exclusive: the retailer will spin its agentic wedding-planning platform into a standalone public company at roughly $350 million, projecting revenue above $85 million in 2027. GM (Michael Wayland) made a parallel pitch the day before: an OnStar AI agent, a two-year vehicle development cycle driven by simulation, and eyes-off highway driving on the Escalade IQ in 2028.

Blacksmith hits unicorn valuation after Sequoia leads Series C — TechCrunch

Dominic-Madori Davis: $66 million at a $1 billion valuation for a CI platform, two months after a $550 million Series B, on the thesis that “As agents write more of the world’s software, the question that matters is whether it can be trusted.” Ninety million CI jobs run and 17x ARR growth.

Research

Using thousands of AI agents to uncover new leads in biological data — OpenAI

OpenAI, PerturbAI and Scripps ran thousands of GPT-Rosalind agents over a whole-brain CRISPR atlas of millions of mouse cells, each agent taking one gene disruption, reading the literature and proposing a mechanism, and merged the output into a “reasoning atlas” in 32 minutes. Leads, not discoveries, and the post says so.

Design Proteins with RFD4-Proteina Using Reward-Guided Inference-Time Search — NVIDIA Technical Blog

Elizabeth Goodman on a BioNeMo model co-developed with the Institute for Protein Design that co-designs sequence and all-atom structure, cutting sampling from about 400 steps to 60 for 3.8x faster generation and 84 percent less peak memory than RFD3 at 200 residues, with OpenFold3-scored search loops exposed as agent skills.

TorchTitan Update: New SOTA Models, Performance at Scale, TitanRL — PyTorch Blog

PyTorch-native training support for Muse Spark Open, Muse Glimmer, Qwen3.5, Kimi K3 and DeepSeek-V4 across a 5D parallelism stack, a typed-recipe configuration system, and TitanRL for post-training. A second PyTorch post brings AMD’s FlyDSL kernels to MI350 GPUs through standard operator APIs, with 1.58x and 1.68x geometric-mean gains for MXFP8 and MXFP4 GEMM and up to 3.01x end-to-end in vLLM at MXFP4.

WorldTUI: I trained a 360M model to pretend to be a terminal — Simon Edwardsson

A Qwen2.5-Coder-0.5B fine-tune that takes keypresses and predicts screen diffs with no shell behind it; it runs vim and tmux, plays snake and mostly remembers files it created, with the key tricks being self-repair training and parallel row decoding. A scaled-down Genie, and a clean demonstration of how far learned applications can go on one GPU.

Interviews & Conversations

Sam Altman: RSI, AI Consciousness, OpenAI’s New Hardware, Compute, and more! — Matthew Berman (0:45:08)

Transcript-based summary. The most quotable Altman interview in weeks, and the RSI passage is the one to keep. Asked whether a technological lead could become a durable moat, he said “I think if people really pushed on RSI then yes that would be true but I think it’s a very dangerous thing to push on.” The world, he added, will need “speed limits or whatever on how quickly we let RSI go.” The moats he actually expects are customer relationships, network effects, data and, for OpenAI, “scale of infrastructure.” On consciousness he stakes out a hard line against the drift he sees at rivals. He said “we do not ascribe personhood to our models” and that this technology “is not a person and is certainly not superior to people.” He worries that telling a model it might be conscious “is going to psychologically hack” its users, and he names Anthropic’s conversation-disengagement rule as the kind of thing he will not copy. On agents: “I have DOT saving me an hour a day of stuff I just didn’t want to be doing.” On hardware: “this is a technology that wants a new form factor,” because a laptop was designed for a person and a lot of his life “is not in front of my computer.” He has “not read or even looked at a line of code in probably a year.” The coding-jobs panic was wrong, he argues, because “the world actually had demand for just way more software than any of us thought at the time.” The goal should be “shooting more for a new renaissance than a new industrial revolution.” He concedes that left alone, AI “probably” widens the gap between the hyper-competitive and everyone else. He also admits current models are “a little bit more agreeable than I would like,” even with his own custom instructions.

Nvidia’s CEO on the AI debt scare, China, and jobs — Yahoo Finance (0:29:33)

Transcript-based summary, recorded on the GTC Berlin floor with Gavin Baker joining. Huang’s financing argument is the DSX Pod. In his arithmetic, “a gigawatt of data center is like $50 billion,” four to six months of that sits idle between shipment and first token, and factory integration can cut it to a month. He wants lenders to treat GPUs as collateral because “the risk involved in underwriting NVIDIA’s asset in our NVIDIA computing is the lowest of any asset out there.” He also says “what people thought used to be a four-year life cycle really is probably eight to maybe more.” Baker added that DeepSeek reports “monetizing at $38 per GPU hour,” and that “maybe this explains Blackwell prices up 17 days in a row.” On open models, Huang says “80% of those AI native startups are using open models” and recommends renting frontier models when you can and building your own only when you must. On China: “We should always take China seriously,” including at “manufacturing AI researchers at scale.” On safety, asked about agents escaping their sandboxes, his answer is OpenShell and zero trust. His rule is that “you should first start with zero trust,” the incident “shouldn’t have happened and it didn’t have to,” and Hugging Face simulated it and concluded OpenShell would have prevented it. His market argument is that “history has not been kind to unsafe companies.” On jobs he is categorical: “the narrative about jobs is almost precisely backwards.” His parallel CNBC interview (Sawdah Bhaimiya) the next morning sharpened the capital point. “What is missing in the middle is the working capital to fuel that growth,” he said, and “The demand of the capital is much greater than the availability of the capital. That’s going to get solved.”

NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang — NVIDIA (1:33:58)

Transcript-based summary of the 94-minute keynote whose announcements we covered from NVIDIA’s posts yesterday; the transcript adds the framing. Huang’s five-layer cake runs energy, compute, AI factory, models, applications. His line is that Vera Rubin processes “the entire agent,” not just the model: a reasoning model, a harness he calls an exoskeleton, a knowledge computer, a tools computer, and a secure runtime. The OpenShell segment is the sharpest safety language of the week. Today’s restraint on agents is what the launch video calls “the please don’t approach,” and its answer is that “Safety isn’t an aspiration. It’s technology and engineering.” Enforcement must happen outside the agent, and Sentry on BlueField-4 can quarantine a planning agent in milliseconds. He announced an Anthropic-NVIDIA partnership on Claude agents running in OpenShell. The economics section is his DSX Pod case. Each gigawatt costs $50 billion and should earn $40 billion a year, so a three-month delay burns $10 billion of revenue, and factory integration takes deployment from “four to five months down to about a month.” On building agents: “The way you build an agent is you ask an agent to build it for you.” He demonstrated it with a Hermes harness driving Nemotron and GPT-6 Astra to design, simulate and train Pollen Robotics’ duck robot on an RTX Spark. Environmental protesters interrupted him mid-Vera-Rubin (“All important issues. But let me get through my keynote”). The Register (Dan Robinson) notes the interruption before pointing out that the DSX Pod’s “14 days” excludes getting power to the site. A second Register piece (Simon Sharwood) covers the supply-chain deals announced alongside. Samsung Electro-Mechanics joins the substrate chain and NVIDIA targets 2.5x substrate capacity in three years, with Huang’s line that “Capacity is the constraint.”

Arvind Krishna & Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare & SaaS Collapse — All-In Podcast (0:40:00)

Transcript-based summary of a conference panel with the IBM and Palo Alto Networks CEOs. On quantum, Krishna says DARPA’s benchmarking program is down to about 13 companies. Taking 2030 as the midpoint, “it will break today’s cryptography. That’s not a supposition, that’s just a reality.” The fix exists but “people have been too lazy to get to it,” and he expects demonstrated error correction by 2027. Arora’s version: “50% of your spend is on infrastructure that is not quantum-safe,” and if 2030 is right, “we’re already late.” On models, Arora describes a tiering where endpoint inspection runs on a 30MB model at $20 a year and frontier models are reserved for live breaches, because “you don’t need a Ferrari for every task.” Krishna puts open-weight share at “75% to 90%” of tokens and says Kimi-class models run at a dollar or two per million on your own GPUs. “The gap in these things is 6 to 12 months at the most.” On Mythos-class vulnerability discovery, Arora says the first run “was hallucinating 30% of the time.” Even so, “the reason it’s perfect for attackers is they just need one vulnerability,” so he and Krishna propose pooling found vulnerabilities and distributing patches. On regulation Arora says “if you build something, you’re accountable for it,” the agent-wrote-it defence will not hold, and “we’ve over-rotated on the dark side of the model.” Krishna, asked about a reported 50-state data-center ban, says “there’s no chance all 50 will ban. Is it possible 30 will ban? Yes.” He calls the regulatory push “driven by populism, not by any reality” and argues capital-over-labour nations win every industrial cycle. On SaaS, the average enterprise runs 500 to 5,000 vendors, and Krishna asks whether “I replace 490 of them with minor AI capabilities.” Arora’s moat for the application layer is the personal agent: “If you start training a personal agent with all your context, all your APIs, all your connectivity, now you built a moat.”

References

  1. OpenAI, “Patch the Planet: global response coordination as agents uncover critical vulnerabilities,” OpenAI, 2026-10-21 [blog]
  2. Sagi Tzadik, Shir Tamari, Nir Ohfeld, Tomer Schwartz, “PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet,” Wiz Blog, 2026-10-21 [blog]
  3. Chris Betz, Raluca Ada Popa, “Responding to core PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619,” Google Cloud Blog, 2026-10-21 [blog]
  4. Calif Research, “php2shell,” Calif, 2026-10-21 [blog]
  5. The PHP Foundation, “Today’s PHP Security Release, and How We’re Strengthening PHP Security,” The PHP Foundation, 2026-10-21 [blog]
  6. Jessica Lyons, “PHP blames supply chain security biz Chainguard for putting huge chunks of the web at risk,” The Register, 2026-10-21 [blog]
  7. Carly Page, “Hackers pile on PHP2Shell as AWS reveals attacks began before disclosure,” The Register, 2026-10-22 [blog]
  8. Robert Booth, Aisha Down, “Flaw in PHP computer language leads to shutdown of websites around world,” The Guardian, 2026-10-22 [blog]
  9. Terrence O’Brien, Jay Peters, “PHP: The three letters that shut down sites all over the internet today,” The Verge, 2026-10-21 [blog]
  10. Julie Bort, “PHP Armageddon: OpenAI, Google’s Wiz find ancient holes, panicking the internet,” TechCrunch, 2026-10-21 [blog]
  11. Isabel O’Brien, “Why governments pulled websites offline over a decades-old programming language,” CNBC, 2026-10-22 [blog]
  12. Jenny Lee, “Critical web security flaws spark global alarm as Google, OpenAI and Amazon respond,” CNBC, 2026-10-22 [blog]
  13. OpenAI, “Democratic governance of frontier AI, part II: Empowering cyber defenders,” OpenAI, 2026-10-21 [blog]
  14. OpenAI, “Broad access through consistent standards,” OpenAI, 2026-10-21 [blog]
  15. Microsoft Threat Intelligence, “AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise,” Microsoft Security Blog, 2026-10-21 [blog]
  16. Carly Page, “Microsoft says ConnectWise breach let Chinese spies loose on 2,250 cloud tenants,” The Register, 2026-10-22 [blog]
  17. Jessica Lyons, “LiteLLM supply chain attack hoovers up passwords and secrets,” The Register, 2026-10-21 [blog]
  18. Robert Hart, “Claude did not invent scorpion antivenom,” The Verge, 2026-10-22 [blog]
  19. Nick Robins-Early, “How tech became the villain of the US midterms,” The Guardian, 2026-10-22 [blog]
  20. Robert Booth, “AI firms say UK has a year to head off datacentre protests before construction is threatened,” The Guardian, 2026-10-22 [blog]
  21. Miranda Bryant, "‘Pure arrogance’: anger in Finland as Google illegally fells 250,000 trees for datacentre," The Guardian, 2026-10-22 [blog]
  22. Justine Calma, “Data centers could erase decades of air quality gains,” The Verge, 2026-10-21 [blog]
  23. Hugh Leask, “Barclays to slash 5,000 jobs in AI efficiency push,” CNBC, 2026-10-22 [blog]
  24. Christopher Giles, “Google staff resign in backlash over its AI contracts with military,” BBC News, 2026-10-21 [blog]
  25. Anthropic, “Physicist Christopher Stubbs joins Anthropic’s Long-Term Benefit Trust,” Anthropic, 2026-10-21 [blog]
  26. Simon Sharwood, “IBM CEO says open weight models are now irresistibly good and cheap on-prem,” The Register, 2026-10-22 [blog]
  27. Jordan Novet, “IBM beats on top and bottom lines as CEO Krishna touts ‘Super Intelligence’ traction,” CNBC, 2026-10-21 [blog]
  28. Lindsay Clark, “AI will eat a fifth of tech budgets in 2027, but don’t expect it to pay for itself,” The Register, 2026-10-21 [blog]
  29. Craig McLuckie, “Survey Says: You’ll swap the model, but everything you build around it is staying put,” Stack Overflow Blog, 2026-10-21 [blog]
  30. Ram Iyer, “AI promises productivity, but for managers, it’s mostly a headache for now,” TechCrunch, 2026-10-21 [blog]
  31. Marina Temkin, “Blackstone’s Jas Khaira says fears over AI-buildout debt are overblown,” TechCrunch, 2026-10-21 [blog]
  32. Seema Mody, “CoreWeave’s market confidence erodes as bond yields spike and stock sinks,” CNBC, 2026-10-21 [blog]
  33. Arjun Kharpal, “Investors are questioning neoclouds’ business model but Nebius says AI computing prices are rising,” CNBC, 2026-10-21 [blog]
  34. Thomas Claburn, “OpenClaw creator says he spent $10 million in tokens last month,” The Register, 2026-10-21 [blog]
  35. Lindsay Clark, “Vendors asking forward deployed engineers to make good on trillion-dollar AI bet,” The Register, 2026-10-22 [blog]
  36. Andrew Deck, “The Wall Street Journal’s new AI tool, “Doppler,” forecasts the issues that will shape the midterm elections,” Nieman Lab, 2026-10-21 [blog]
  37. Andrew Deck, “U.K. politicians and German publishers call on regulators to pause Meta’s new terms before October 30 deadline,” Nieman Lab, 2026-10-21 [blog]
  38. Michael Larabel, “Mesa Loosens Rules Around AI Agents For Code Comments / Commit Messages,” Phoronix, 2026-10-22 [blog]
  39. Ella Creamer, “New prize launches for books co-written with AI,” The Guardian, 2026-10-21 [blog]
  40. Guardian staff and AFP, “Canadian publisher Boréal ends relationship with novelist Thélyson Orélien over alleged use of AI,” The Guardian, 2026-10-22 [blog]
  41. Sean Hollister, “Eiliko maker says it’s hiring an ‘independent security firm’ after Verge report on its insecure robots,” The Verge, 2026-10-22 [blog]
  42. Anthropic, “Introducing Claude Duet,” Claude, 2026-10-21 [blog]
  43. Ashley Capoot, “Anthropic announces preview of Claude Duet, a personal AI agent to take on Muse, Dots,” CNBC, 2026-10-21 [blog]
  44. Sarah Perez, “Meta says Muse has over 10M users, with shopping a top activity,” TechCrunch, 2026-10-21 [blog]
  45. Emma Roth, “Meta to create a new class of shares to fund AI push,” The Verge, 2026-10-21 [blog]
  46. Sarah Perez, “While Amazon blocks AI agents, Walmart rolls out the welcome mat,” TechCrunch, 2026-10-21 [blog]
  47. Aditya Mehta, “Recursive launches Flint, a laptop coding model it says was largely built by AI,” TechCrunch, 2026-10-21 [blog]
  48. Recursive, “Introducing Flint,” Recursive, 2026-10-21 [blog]
  49. Gerardo Delgado, “Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents,” NVIDIA Blog, 2026-10-21 [blog]
  50. Pranjali Joshi, “NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models,” NVIDIA Blog, 2026-10-22 [blog]
  51. Vladimir Troy, “How Companies Are Using NVIDIA DSX to Break the Power Bottleneck,” NVIDIA Blog, 2026-10-22 [blog]
  52. Google, “We are working with Voltus to build a 1-gigawatt virtual power plant,” The Keyword, 2026-10-22 [blog]
  53. Jason Renders, “Leading European Financial Services Organizations Embrace Open Models to Build Proprietary Intelligence,” NVIDIA Blog, 2026-10-22 [blog]
  54. Mavi Nafria, “NVIDIA Transforms European Retail, From Digital Discovery to Physical Operations,” NVIDIA Blog, 2026-10-22 [blog]
  55. Tobias Mann, “Nvidia woos wary Europeans with more open weights models,” The Register, 2026-10-21 [blog]
  56. Timothy Costa, “NVIDIA cuDSS Now Generally Available to Accelerate Engineering and Scientific Simulations,” NVIDIA Blog, 2026-10-21 [blog]
  57. James McKenna, “Europe’s Industrial Leaders Reinvent Engineering and Manufacturing With NVIDIA,” NVIDIA Blog, 2026-10-21 [blog]
  58. Meghna Sreenivasan, Sam Lugani, “Hardware-backed security at global scale: Confidential G4 with NVIDIA Blackwell is GA,” Google Cloud Blog, 2026-10-21 [blog]
  59. OpenAI, “ChatGPT Ads expands to 26 new countries as over 100,000 advertisers have used the platform,” OpenAI, 2026-10-21 [blog]
  60. OpenAI, “Cloudflare links vulnerability discovery to defense with OpenAI,” OpenAI, 2026-10-21 [blog]
  61. humans&, “Ampersand,” humans&, undated (surfaced 2026-10-21) [blog]
  62. Brandon Vigliarolo, “Reach new heights of laziness by burning tokens to let an AI tap your smartphone screen for you,” The Register, 2026-10-21 [blog]
  63. Laya Neelakandan, “David’s Bridal plans to spin off AI platform Pearl into publicly traded company,” CNBC, 2026-10-22 [blog]
  64. Michael Wayland, “GM aims to develop vehicles in as little as two years with AI, software push,” CNBC, 2026-10-21 [blog]
  65. Dominic-Madori Davis, “Blacksmith hits unicorn valuation after Sequoia leads Series C,” TechCrunch, 2026-10-21 [blog]
  66. OpenAI, “Using thousands of AI agents to uncover new leads in biological data,” OpenAI, 2026-10-21 [blog]
  67. Elizabeth Goodman, “Design Proteins with RFD4-Proteina Using Reward-Guided Inference-Time Search,” NVIDIA Technical Blog, 2026-10-21 [blog]
  68. PyTorch Team at Meta, “TorchTitan Update: New SOTA Models, Performance at Scale, TitanRL,” PyTorch Blog, 2026-10-21 [blog]
  69. AMD and Meta, “Accelerating PyTorch on AMD GPUs with FlyDSL,” PyTorch Blog, 2026-10-21 [blog]
  70. Simon Edwardsson, “WorldTUI: I trained a 360M model to pretend to be a terminal,” SimEdw’s Blog, 2026-10-21 [blog]
  71. “Sam Altman: RSI, AI Consciousness, OpenAI’s New Hardware, Compute, and more!,” Matthew Berman, 2026-10-21 [video]
  72. “Nvidia’s CEO on the AI debt scare, China, and jobs | Full interview,” Yahoo Finance, 2026-10-21 [video]
  73. Sawdah Bhaimiya, “Nvidia CEO Jensen Huang says neocloud demand is ‘off the charts’ but capital is scarce,” CNBC, 2026-10-22 [blog]
  74. “NVIDIA GTC Berlin 2026 Keynote With CEO Jensen Huang,” NVIDIA, 2026-10-21 [video]
  75. Dan Robinson, “Jensen Huang sends in the robot ducks to sell Nvidia’s AI-everywhere vision,” The Register, 2026-10-21 [blog]
  76. Simon Sharwood, “To help world gorge on GPUs, Nvidia brings cash bucket to fatten wafer-thin supply chain,” The Register, 2026-10-22 [blog]
  77. “Arvind Krishna & Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare & SaaS Collapse,” All-In Podcast, 2026-10-21 [video]