Claude Duet, Walmart's Agent Shop, and a PHP Scramble

Grant: Anthropic built an AI agent that checks in on you instead of waiting around for a prompt.

Claire: Walmart just opened its doors to AI shopping agents while Amazon is slamming them shut.

Grant: And two AI models found a decade-old flaw sitting in a huge chunk of the web, which means somebody's site probably needs patching right now.

Claire: I'm Claire.

Grant: I'm Grant, and this is The Week in AI from The Z Media Podcast, covering the week ending Friday, October 23, 2026.

Grant: So Claire, Anthropic put out something new this week called Claude Duet. The Verge and The Decoder both covered it. It's a research preview, built on Claude Opus 5.5, and it's live on iOS, Android, and the Claude web app.

Claire: Research preview meaning not everyone can touch it yet?

Grant: Right, it's waitlist only, and only for Pro and Max subscribers in the US. Team and Enterprise aren't included yet. The idea is it works in the background across your connected apps and then reaches out to you, instead of you opening it up and typing a question.

Claire: What's it actually doing while it's back there? Because that's the part that makes business owners nervous.

Grant: It sends a short morning summary pulling from your connected apps, then only pings you again when something needs attention. One ongoing thread, you can talk to it by text or voice. It's got its own memory and skills carried over from your Claude account, and it reaches tools like Gmail or Shopify through Anthropic's connector marketplace.

Claire: Okay, but can it spend my money or send an email without asking?

Grant: No. It has to show you its plan and get approval before it pays, signs, messages, or deletes anything. And here's the one I liked, moving money between accounts is blocked even if you approve it. Anthropic built in a hard stop there.

Claire: So what should a small business owner actually hand off to it first?

Grant: Think scheduling, pulling dashboards together, forecasting, chasing invoices. The boring recurring stuff. Anthropic is upfront that it can misread a thread or miss something urgent, and it only checks your accounts every few hours because background checks count against your subscription limits. So it's not instant, and it's not for anything time sensitive yet.

Claire: That's a fair tradeoff for a first version. Watch and see how it fits next to regular Claude.

Claire: Next one's mine. OpenAI is widening ChatGPT advertising to 26 more countries, according to The Information, and the company says it now has more than a hundred thousand advertisers.

Grant: A hundred thousand advertisers sounds like a real ad business already.

Claire: That's the headline number, but MarTech's own framing on this is blunt. Plenty of advertisers, not much revenue. So wide participation, but the dollars per advertiser look thin so far.

Grant: Do we know anything about formats or pricing, what it actually looks like inside a chat?

Claire: Not from what's available to us. The source articles on specifics weren't fully accessible, so I can't give you targeting details or cost per click.

Grant: So for a marketer sitting in the newly added countries, is this worth testing now or waiting?

Claire: I'd call it a small, well-tracked test, not a budget shift. A big advertiser pool can also mean more competition for the same placements down the road, so getting in early isn't nothing. But modest revenue signals mean spend per advertiser is still small. Watch for OpenAI to publish real targeting and measurement details before you commit real money.

Grant: Claire, this one's interesting even if you're nowhere near pharma. OpenAI acquired a company called Orchestra Bio. TechCrunch, The Information, and The Decoder all picked it up.

Claire: What does Orchestra actually do?

Grant: Four-year-old startup, builds AI agents for the operational side of drug development. Think tracking clinical trials, flagging which timelines get disrupted if something goes wrong, like a failed production batch, and surfacing next steps. TechCrunch reports they'd raised around twenty million dollars as of 2025, backed by Andreessen Horowitz, Atria, and 8VC. Deal terms weren't disclosed.

Claire: And OpenAI's folding that into something they already had?

Grant: Yes, into Rosalind, their life sciences model, and the Rosalind Workbench tool. OpenAI says Moderna and Novo Nordisk are already Rosalind customers, so this isn't a side project, it's built for buyers with large, complicated operations.

Claire: Why should a marketing agency care about a pharma deal?

Grant: Because it's a pattern, not a one-off. TechCrunch frames it as a startup being absorbed rather than outcompeted by a lab. The big labs aren't just building general models anymore, they're buying their way into owning entire workflows inside an industry.

Claire: So if you've got clients in those spaces, you watch who the platform owner becomes.

Grant: Exactly, and you plan your partnerships and your pitch around that.

Claire: Okay, my turn again. Meta released something called Muse Spark Open, and the surprising part isn't the model, it's what it runs on. The Verge, TechCrunch, and Wired all covered this.

Grant: Runs on, meaning what, a laptop?

Claire: Not quite that casual. It's open-weight, 816 billion parameters, a one million token context window, and Meta says it runs on two Apple M5 Ultra Mac Studios using Apple's MLX framework. TechCrunch, citing Engadget, says you'd need at least one terabyte of RAM. So this is desktop hardware, but it's the high end of desktop hardware.

Grant: Wait, Meta and Apple worked together on this? Those two aren't usually in the same sentence.

Claire: They are here. The Verge, citing Axios, says the companies worked together for months, and Meta's Mark Zuckerberg and Apple's John Ternus publicly praised each other's work. Apple even published its own technical post about MLX and the model.

Grant: What's the actual pitch for a content team here? Why would anyone want to run a model locally instead of just calling an API?

Claire: Cost efficiency over time, more transparency since it's open-weight, and keeping sensitive campaign or client material on your own machines instead of sending it out to a cloud service. Meta says it tested the model doing agentic work, fixing code in a Mastodon client, reviewing and cropping footage in DaVinci Resolve.

Grant: Those are Meta's own claims though, not independent testing.

Claire: Correct, and that's the honest caveat. No independent benchmarks yet. Realistically this is for agencies and larger teams who can afford the hardware. Everyone else can test it through one of the API platforms first before deciding if the investment makes sense.

Grant: Alright, this next one isn't glamorous, but if your website runs on PHP, which a huge share of the web does, you need to hear it. The Information, The Decoder, and The Rundown AI all reported on it.

Claire: PHP as in the language behind a ton of content management systems.

Grant: That one. Two critical, decade-old flaws were found. OpenAI's GPT-5.6 Sol found them through something called Patch-the-Planet, and separately, security firm Wiz found the same flaws on its own using Sol alongside Google's Gemini 4 Argon.

Claire: How bad are we talking?

Grant: The Rundown AI reports these flaws let attackers take over a server without even logging in. Fixes got coordinated over two months, in five phases, with PHP developers, Wiz, Cloudflare, CISA, and the White House all involved. The NSA reportedly urged anyone who couldn't patch immediately to just take their systems offline, and Cloudflare said it was already seeing early attacks.

Claire: That's not a next-quarter fix, that's a this-week fix.

Grant: Right, and there's a GitHub tool server operators can use to check if their PHP version is affected. We'll put the links in the show notes. For marketing teams, your blog, your landing pages, your CMS, those are probably sitting on PHP somewhere, often managed by a host or an agency you never think about.

Claire: So the actual action item is just asking the question.

Grant: Exactly, call your host or your developer this week and ask directly, has the patch been applied. A compromised server means defacement, lost leads, tampered tracking, and a dent in brand trust you don't want.

Claire: This next story is a genuinely interesting split between two retail giants. Walmart just opened its site to AI shopping agents. The Verge, TechCrunch, and Adweek all covered it.

Grant: Meaning an AI can just go buy stuff on Walmart for you now?

Claire: If it's approved, yes. Walmart launched something called Agent Gateway, built on the Model Context Protocol, letting approved agents search inventory and complete purchases, including at Sam's Club. Walmart's CEO John Furner talked about it at the Masters of Scale Summit.

Grant: How does a shopper keep an agent from going rogue with their credit card?

Claire: Agent Profiles, rolling out next month. Each agent gets its own profile, its purchases show up separate from your own, approval is required by default, or you can set a spending cap instead, like a hundred dollars. Walmart compared it to adding an authorized user to a credit card.

Grant: And if the agent misbehaves?

Claire: It gets blocked if it breaks Walmart's terms, and the normal returns process still applies even if an agent buys the wrong thing. Meanwhile, Amazon has actually blocked agents like Perplexity's shopping tool and Meta's Muse assistant. Walmart's doing the opposite, it already works with Muse and offers integrations with ChatGPT and Gemini.

Grant: So what does this mean for a brand that sells through Walmart or competes for shelf space there?

Claire: If agents start doing more of the browsing and choosing, your product data, your availability, and your pricing start mattering more than clever on-page persuasion. Watch how many agents Walmart actually approves, and whether other retailers follow this open approach or stay closed like Amazon.

Grant: Speaking of Muse, Claire, it's had a wild couple of weeks. TechCrunch and The Verge both covered this one.

Claire: Give me the growth numbers first.

Grant: Sensor Tower counts 10.9 million downloads of Muse since its September debut, 10.2 million of those in the US, in just 40 days. Only 11 other apps have ever hit that mark that fast, mostly mobile games, plus Threads and Disney+. Meta's backing it hard too, Muse took 39 percent of Meta's ad spend in October, up from 23 percent in September.

Claire: Are people actually sticking around, or is that just download hype?

Grant: Retention looks solid. 30 percent at 7 days, 23 percent at 14 days, both ahead of the average for rivals like ChatGPT, Claude, and Grok. Daily active users hit 2.5 million on October 16. It's US and Canada only right now, so it's not a global comparison yet.

Claire: You said pause though. What went wrong?

Grant: Meta paused Muse's conversation-sharing feature. The Verge, citing Ars Technica, reports a security firm called Claroc found more than 6,800 shared chat links just sitting out there through public DNS records. Each link's address was built from the chat's title plus a random ID, which apparently made them easier to stumble onto than anyone intended.

Claire: What kind of chats are we talking about here?

Grant: Claroc says it saw a marketing company's revenue review, a loan application, and a staff calendar, just out there exposed. And this isn't only a Muse problem, similar sharing features on ChatGPT, Claude, and Grok have also surfaced chats in Google Search before.

Claire: So the lesson for a marketing team isn't just about Muse, it's about anything you paste into any AI chat.

Grant: Right. A fast-growing, well-retained assistant is worth watching for audience planning, but don't put revenue numbers, client data, or strategy docs into these tools assuming they're private. Watch how Meta changes link discovery before sharing comes back.

Claire: Last one's mine, and it's a cost story. Deepseek shipped a new model called V4.1-Flash-1017. The Decoder and MarkTechPost both covered it.

Grant: Cost story meaning it's cheap to run?

Claire: That's the pitch. It's open-weight, available on Hugging Face and ModelScope, plus Deepseek's own app, website, and API. On their own benchmark, DeepSWE, they report 77 percent accuracy at just under twenty cents per task. Their prior version scored 74.2 percent at around fifty cents. A lower setting still tops 72 percent for roughly ten cents.

Grant: How does that stack against the big names?

Claire: Deepseek's own chart puts rivals like Claude Opus 5, GPT-6 Astra, and Gemini 3.8 Flash anywhere from two dollars to over ten dollars per task, with lower accuracy than their new model claims. But I want to be clear, every single one of these numbers is Deepseek's own reporting. No independent benchmarks exist yet.

Grant: You said search API too, what's that about?

Claire: They're opening their internal search engine, which they say handles hundreds of millions of queries a day, through Search and Fetch APIs at seventy five cents per thousand calls, per endpoint. And through October 31st, you can try it free inside Deepseek Harness, Claude Code, or Codex.

Grant: So for a content team building research or monitoring agents, this could just lower the bill.

Claire: Potentially, yes, if the real-world quality matches the benchmark. Benchmarks and lived experience don't always line up. Worth watching for independent evaluations, and also worth knowing China is reportedly weighing tighter rules on releasing open-weight models, which could affect how comfortable teams feel depending on these in production.

Grant: Alright Claire, if a marketing team only does one thing next week, what is it?

Claire: Easy one. Go ask whoever manages your website whether the PHP patch has been applied.

Claire: Do that this week.

Grant: That's the week. Agents that check in on you, agents that shop for you, and a security scramble that's a good reminder that even boring infrastructure needs attention.

Claire: If all of this makes you wish your own content and customer systems just ran themselves safely in the background, that's exactly what we build. Z Media puts together fully automated AI content and CRM systems that write, design, film, publish, and even answer customers for you.

Grant: Talk with our team, the links are in the show notes, or call or WhatsApp us at four one nine, five zero zero, four eight eight two. And if you want the key points every single day, not just on Fridays, subscribe to the AI news wire.

Claire: Thanks for spending part of your week with us. We'll see you next Friday.