AI News Today — October 22, 2026: Muse Tops 10 Million Users as Shopping Takes Off, Claude Duet Arrives, and Walmart Opens to AI Agents

The big picture

The agent race stopped being a demo contest and became a shopping and work contest. Meta used its earnings call to say Muse has passed 10 million users and that people are already trusting it to buy things. Anthropic launched its own proactive agent, Claude Duet, aimed at professionals. And Walmart built a formal front door for agents from ChatGPT, Gemini, and Muse — while Amazon is pushing agents out.

Underneath the product news, two warnings landed. A pair of PHP flaws hidden since 2004 forced an urgent patch-or-go-offline warning, and a compromised release of a popular AI routing library showed how much damage one bad package can do. Agents are getting real access. The controls around them now matter as much as the models.

What happened today

Muse AI: Meta confirms more than 10 million users — and shopping is already a top activity

What happened: On its third-quarter earnings call on October 21, Meta said Muse has been adopted by more than 10 million users, confirming the Sensor Tower estimates reported earlier this week. CEO Mark Zuckerberg said about 25% of people who set up Muse now use it daily, and that over a quarter of Muse's engaged users who were shopping via the agent had an average purchase size of over $200. He quoted payments firm Plaid — which provides the connector Muse uses for agentic shopping — as saying it had not seen activity like this on any other service before. Beyond shopping, top uses include personal productivity tasks such as connecting Muse to email and calendars. Zuckerberg pointed to Muse Secure VM as part of why users trust the setup, and said a forthcoming Muse Confidential VM will prevent even Meta from seeing users' content. He said Muse has reached product-market fit, that Meta will now focus on scaling it and improving its compute cost efficiency by more than 10 times in the next year, and he teased a future "Muse network" where one person's Muse could work with friends' and families' Muses and with businesses' agents.

Who it matters to: Anyone building for agent commerce or personal agents. Downloads alone were yesterday's story. Today's numbers are about daily use and money moving — shoppers spending over $200 on average through an agent is a behavior change, not a curiosity click.

What you can do: If you sell online, assume customers will soon arrive through an agent, not just a browser. Check how your products appear to agents — clear prices, clear return terms, and machine-readable details — before you need the traffic. If you use Muse, keep using its approval controls for purchases, exactly as you would for any new assistant with spending power.

Learn more: TechCrunch on Meta confirming Muse has over 10M users, with shopping a top activity and TechCrunch on Muse hitting 10M U.S. downloads in 40 days

Gemini Spark: no new launch — Task mode and the November 17 deadline still stand

What happened: There was no new Gemini Spark announcement on October 21–22. The live facts are unchanged from earlier this month: Google is folding Spark into the main Gemini app as Task mode, with a Chat/Task switch above the prompt box and a new Scheduled page, in a server-side Android rollout first reported by 9to5Google on October 12. Spark still requires a Google AI Pro or AI Ultra subscription. The separate deadline Gemini users should keep on their calendars is also unchanged: Gems are becoming Skills, and personal-account Gems are set to auto-migrate on November 17, 2026.

Who it matters to: Gemini users who rely on Gems for repeat work. The name on the button keeps changing. The migration date is the part that can break a workflow.

What you can do: Save a copy of each important Gem's instructions and files somewhere you control, and note any Gem that depends on tools or files Skills may not support yet. Do not wait until the week of November 17 to check what moved.

Learn more: 9to5Google on the Gemini 'Task' redesign integrating Spark

Anthropic launched Claude Duet, a proactive agent that comes to you

What happened: Anthropic announced Claude Duet on October 21 as a research preview for Pro and Max subscribers in the United States, rolling out gradually through a waitlist. Duet is a standalone app for iOS and Android and on the web at claude.ai, powered by Claude Opus 5.5. Instead of waiting for a prompt, it starts the day with a short brief drawn from your connected email, calendar, and work apps, then checks those apps every few hours and messages you only when something needs attention. It works on its own cloud computer with its own browser, so it keeps working after you close the app, and it carries over the same memories, skills, and connectors as regular Claude. Anthropic says Duet is designed to ask before it sends a message, makes a payment, signs a document, or deletes anything. Larger jobs come back as files — documents, spreadsheets, decks, and dashboards — in a separate tab. Duet is included at no additional cost, but all of its work, including background checks, counts toward your existing plan usage limits. It is not yet available on Team or Enterprise plans.

Who it matters to: Professionals and small-business owners deciding between the three new proactive agents — Muse, OpenAI Dots, and now Duet. Duet is the one framed most clearly around business admin: inbox, calendar, invoices, and follow-ups.

What you can do: If you are a U.S. Pro or Max subscriber, join the waitlist and start with one low-risk job, such as a morning brief or a follow-up tracker. Watch your usage meter for the first week — background work spending your limits is the cost to understand before you rely on it.

Learn more: 9to5Mac on Anthropic taking on ChatGPT Dots and Meta Muse with Claude Duet and MacRumors on the Claude Duet launch

Walmart built a gateway for AI agents while Amazon blocks them

What happened: Walmart announced an Agent Gateway on October 21 — its own MCP server that lets approved AI agents connect to Walmart and Sam's Club to search inventory and complete purchases. Walmart already works with Muse and offers integrations with ChatGPT, Gemini, and other assistants. Starting next month, Agent Profiles will let each customer authorize individual agents inside their Walmart account, see agent purchases separately from their own orders, and either require approval for each transaction or set a spending limit — a setup Walmart compared to adding an authorized user to a credit card. Any Walmart account holder can use it; a Walmart+ membership is not required. Walmart also set terms for agents: they cannot access customers' payment card data, cannot act beyond what the shopper intends, and agents that break the rules will be blocked. Its normal returns process applies even when an agent makes the purchase. The contrast is deliberate and recent — Amazon has begun blocking third-party AI agents from buying on behalf of shoppers.

Who it matters to: Every online seller and every customer who might shop through an agent. Walmart is betting that controlled access wins more sales than blocking agents wins control. Amazon is betting the opposite. One of those bets will shape how agent commerce works everywhere else.

What you can do: If you shop with an agent, prefer stores that give you per-agent limits, a separate purchase history, and a clear right to revoke access — and set those controls before the first purchase, not after a mistake. If you sell online, copy the checklist: per-agent identity, spending caps, approval rules, and returns that still apply when an agent buys.

Learn more: TechCrunch on Walmart rolling out the welcome mat while Amazon blocks AI agents and Reuters via The Hindu BusinessLine on Walmart's AI shopping tools for ChatGPT, Gemini, and other agents

Two PHP flaws hidden since 2004 triggered a patch-or-go-offline warning

What happened: The U.S. National Security Agency warned on October 21 about two critical flaws in PHP that have existed in all versions since PHP 5 in 2004. PHP powers roughly 70% of websites by one widely cited count, including many WordPress sites. The flaws can let attackers reach a server's file system or a web app's database. Patches have been issued by the PHP Foundation, and the guidance is blunt: update to the current release, PHP 8.5.12 — noting that PHP 8.1 and older are no longer supported — and if a site cannot be patched immediately, it should be taken offline until it can be. The flaws were found with AI help: a security firm working in OpenAI's Patch the Planet initiative used GPT-5.6 Sol to find one flaw, and Google's Wiz used its Atlas security agent, Gemini 4 Argon, and GPT-5.6 Sol to find vulnerabilities as well. Both followed coordinated disclosure and worked with PHP on patches before announcing them, and there were no reports of exploits in the wild before disclosure. Wiz's warning explains the urgency: the flaws were previously too complex to exploit by hand, but AI now lowers that barrier.

Who it matters to: Anyone who runs a website — especially a WordPress site on an old hosting plan — and anyone whose business depends on one. This is not a story about someone else's server. PHP's reach makes it almost everyone's problem.

What you can do: Today, check what PHP version your sites and your clients' sites are running, and ask your host in writing when the patch will be applied. If a host cannot patch promptly, follow the official guidance seriously rather than hoping attackers are slow.

Learn more: TechCrunch on the PHP flaws found by OpenAI and Google's Wiz

Recursive shipped Flint, a free laptop coding model largely built by AI

What happened: Recursive released Flint on October 21, a free coding model designed to run on a laptop. Flint is a mixture-of-experts model with 40 billion parameters, of which only 2.5 billion are active at a time, and it is available to download on Hugging Face under the Apache 2.0 license. Recursive says its automated research system, the Eureka Machine, helped with nearly every step of building Flint — from pretraining to post-training and inference — allowing a small team to build the model from scratch in under four months. In Recursive's own tests, Flint generated about 131 tokens per second on a MacBook Pro with an Apple M4 Pro chip, roughly twice the speed of similar models it compared on the same machine, and topped 150 tokens per second on Nvidia's RTX Spark. On SWE-Bench Verified, Recursive says Flint outperformed other models under 50 billion parameters. The company also acknowledged Flint needs more guidance than the large paid coding assistants, and an early hands-on test reported by TechCrunch found it handled several small programs reliably.

Who it matters to: Developers who want coding help without sending private code to a cloud service or paying per token. A capable model that runs locally, offline, and for free changes the privacy and cost math for solo builders and small teams.

What you can do: If you have a modern laptop with headroom, download Flint and test it on one small, real task you can easily verify — a utility script or a bug fix — broken into small steps. Compare its result and speed against your current assistant on the same task before changing your workflow.

Learn more: TechCrunch on Recursive launching Flint, the laptop coding model largely built by AI

Also worth knowing

• Blackstone says AI-buildout debt fears are overblown. Blackstone's Jas Khaira argued at TechCrunch Disrupt that most data-center debt is ultimately backed by highly creditworthy hyperscalers, even as neocloud debt has ballooned — CoreWeave's alone to about $50 billion by one cited count — and borrowing costs have widened. He conceded returns from AI adoption inside portfolio companies are real but uneven. TechCrunch on Blackstone's position
• Blacksmith became a unicorn on the back of AI-written code. The continuous-integration platform raised a $66 million Series C led by Sequoia at a $1 billion valuation, just two months after its $45 million Series B. The company says revenue grew 17 times and it has run over 90 million CI jobs, as teams struggle to test the growing volume of code agents produce. TechCrunch on Blacksmith's Series C
• Managers say AI is making their jobs harder, even as they back it. In a Nielsen survey of 1,700 directors and senior managers commissioned by Monday.com, more than half said AI makes it harder to evaluate employees' work, nearly half spend more time reviewing output, and 58% expect to oversee more AI agents than human employees within a year — while 94% said they are excited to improve work with AI. TechCrunch on the manager survey
• A popular AI routing library was backdoored on PyPI — again. Version 1.107.0 of LiteLLM, a Python package that routes requests to AI providers, was published on October 21 with a credential stealer and a persistent backdoor, according to reports from SafeDep, Socket, StepSecurity, and others covered by The Hacker News. PyPI quarantined the release and the malicious Docker images were removed after an exposure window of roughly two hours. LiteLLM said attackers used stolen GitHub tokens from three engineers' accounts, and that the tokens have been revoked. Datadog notes the package has about 23 million weekly downloads and that this is its second supply-chain compromise in 2026. If your team installed 1.107.0, follow the incident guidance and rotate credentials on affected machines — uninstalling alone is not enough. Datadog Security Labs on the LiteLLM compromise
• OpenAI's models are arriving inside Google Cloud. From October 21, selected enterprise customers can access OpenAI models including GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna through Model Garden on Google Cloud's Gemini Enterprise Agent Platform in private preview, billed through Google Cloud Marketplace, according to a Google Cloud summary. Treat pricing as unconfirmed until Google publishes its own rates — do not assume they match OpenAI's API prices. Summary of Google Cloud's OpenAI-models announcement

One practical takeaway

Before you give any agent more access, set the Walmart-style controls — even if the app does not offer them yet. Today's stories point to one checklist that works for Muse, Duet, Spark in Task mode, and any shopping agent: give each agent its own identity where you can, set a spending limit or require approval for every purchase, keep its purchases and work in a history you can review separately, and know how to revoke access in one step. Then do the unglamorous half of the same job: patch PHP on every site you control, and check whether any machine installed LiteLLM 1.107.0 on October 21. Agents are only useful when the guardrails and the plumbing underneath them are sound.

---

Sources: TechCrunch AI, Reuters via The Hindu BusinessLine, 9to5Mac, MacRumors, 9to5Google, The Hacker News, Datadog Security Labs. Summaries are original; follow the links for full reporting. Muse user, daily-use, and shopping figures are Meta's, stated on its earnings call as reported by TechCrunch; Flint speed and benchmark claims are Recursive's own tests as reported by TechCrunch; manager figures are from the Nielsen survey for Monday.com as reported by TechCrunch.