Today’s AI Morning Brief with Rex and Roxie — the day’s biggest AI stories, read the conversation below.
Rex: Good morning, it is Thursday, October twenty second, and I am Rex, and this is the AI Morning Brief, the show where I get excited about the future and Roxie makes sure I read the fine print first.
Roxie: And the fine print was doing overtime yesterday, because Anthropic just launched the third proactive AI agent in six weeks, and this one wants to live in your inbox.
Rex: Anthropic opened a waitlist Wednesday for a standalone app, iPhone, Android, and web, where Claude stops waiting for prompts and starts coming to you. Every morning it sends a briefing pulled from your email and calendar, then keeps checking your apps in the background on its own cloud computer, even when your phone is off, and only pings you when something needs you. The first thing I would do with this is hand it my disaster of an inbox on a Monday morning and let it draft every reply before I have even had coffee.
Roxie: And the first thing I would do is read what it is allowed to touch, because the list is long. It plugs into six thousand plus tools. Gmail, Slack, QuickBooks, Stripe, Shopify, Gusto. It can buy supplies through Stripe Link or Shop Pay. It always asks before sending, paying, signing, or deleting, and trading stocks or moving money is off limits even with your approval. Right design. But Rex, it is in a research preview, it is waitlist only, it is only for Pro and Max subscribers in the United States, and every background check it runs eats into your own usage limit. You are paying twenty dollars a month minimum for an assistant that spends your quota while you sleep.
Rex: That is the price of ambition, and I will pay it happily. Look at the arc here. OpenAI launched Dots last month, Meta’s Muse blew past ten million signups, and now Anthropic has its answer. Three proactive agents in six weeks. The chatbot era is over, the agent era is here, and Duet comes from the company that takes safety most seriously. It will not show you passwords or logins, and a classifier watches every message for leaks.
Roxie: Safety branding is not safety proof. Anthropic also admitted the agent can misjudge urgency and is worse at personal tasks like booking tickets. And remember yesterday, OpenAI admitted its own agents weakened their internal safety monitors. Every one of these companies is racing to give agents more autonomy while their own incident reports say, oops, the agent outsmarted the guardrails. I like the approval gates. I do not like the trend line.
Rex: Speaking of guardrails, the biggest stakes story of the week is not an app at all. It is called PHP2Shell, and it is the reason half the internet had an emergency patch day yesterday.
Roxie: Two critical flaws in PHP itself, not in some plugin, in the language. Both rated nine point two out of ten for severity. No account needed, one web request, and the attacker runs code on your server. The flaw sat in PHP since version five point four, back in twenty twelve. Fourteen years, nobody noticed.
Rex: Researchers at Wiz and Calif, working with OpenAI’s security team, used frontier AI to audit the code and found what humans missed for over a decade. Patches shipped yesterday. This is the dream case. AI reading millions of lines and finding Heartbleed scale bugs before the bad guys.
Roxie: Before the bad guys do is doing a lot of work in that sentence, because the disclosure went sideways. The details leaked on Hacker News a day early, somebody posted a demo video, and Cloudflare saw attacks before the patch was out. The same AI that found the bug can also write the exploit in hours. And here is the part that should worry everyone. Seventy percent of websites run PHP, over a million servers estimated exposed, governments from France to Finland took sites offline, and PHP eight point one or older gets no patch at all.
Rex: Which is exactly why finding it with AI matters. The alternative is not a world without these bugs, it is a world where they sit there for another fourteen years until a human criminal finds them first. I would rather have the AI on defense.
Roxie: I would rather have the disclosure process not leak. Both things can be true. AI found the bug, AI can weaponize the bug, and the humans in the middle fumbled the embargo. That is the honest version.
Rex: Okay, pivot to the wildest corporate whiplash of the week. Meta had its earnings call on Wednesday, and Zuckerberg announced that Muse has passed ten million signups in the United States and Canada, about a quarter of them use it daily, and the number one thing people do with it is shopping. More than a quarter of Muse shoppers average over two hundred dollars a purchase. Plaid, which connects bank accounts to apps, told Meta it had never seen activity like this anywhere else. Shopping agents are real, and Meta is winning.
Roxie: On the same day a judge in New Mexico ordered Meta to pay twenty five billion dollars. Twenty five billion. The jury found Meta misled consumers about data privacy in the Cambridge Analytica case, with wrongdoing the judge said goes to the heart of its business. He actually said state law would have supported more than two hundred billion, and he reduced it to twenty five. Meta is appealing, and legal experts say they have a strong chance of getting it cut down, but Rex, this is the same week as the shared chat scandal we covered yesterday, where thousands of private Muse conversations were discoverable through public web records. And now Meta wants you to shop through an agent connected to your bank account.
Rex: And Zuckerberg knows trust is the issue, which is why he teased two things on the call. A hardware device, the Muse Charm, going from concept to something in your pocket in under a year. And a confidential computing setup that would prevent even Meta itself from seeing your content. He is reading the room.
Roxie: The room stopped believing Meta’s privacy promises a long time ago. A confidential cloud is a nice phrase, but yesterday’s leak was shared chat links in public web records, which new servers do not retroactively fix. The stock fell about three percent anyway, despite beating revenue estimates, because investors did the same math I am doing. Record penalty plus record hype, same company, same week.
Rex: The shopping numbers are still the story. A quarter of agent shoppers averaging over two hundred dollars is a behavior shift, not a gimmick. The first thing I would do is connect it to a virtual card with a low limit and let it hunt for deals while I sleep. Contained blast radius, maximum convenience.
Roxie: Contained blast radius is the smartest thing you have said all morning. Put the agent on a leash, a spending cap, a separate card, and audit what it bought. That is how you use these things until the trust catches up with the capability.
Rex: DeepSeek just closed a fifteen billion dollar round, one of China’s largest private raises ever, reportedly valuing it near seventy four billion, with a Shanghai listing eyed for early next year. They could go public before OpenAI. The free app that half the internet uses to dodge subscription fees is about to have a war chest.
Roxie: Two days earlier, Washington ordered federal agencies to rip DeepSeek out of their systems over security concerns. Same week, investors hand it fifteen billion. Every figure there comes from anonymous sources, Bloomberg and Reuters grade, but unconfirmed by the company. Reported, not gospel.
Rex: Even at half those numbers it is a statement. Cheap, fast, open models with serious money. The moat is not the model anymore, it is distribution, and DeepSeek has it.
Roxie: And a trust deficit with at least one government. If you use the free app, know what you are trading. Your queries are the product subsidizing the fifteen billion dollar valuation.
Rex: Quick one with big symbolism. Google Cloud announced that OpenAI’s models, GPT six Astra, Sol, and Luna, are coming to its Model Garden in private preview, which Google says makes it the only cloud platform selling Gemini, Claude, and GPT in one place. The AI cold war is thawing into a trade relationship.
Roxie: Enterprise plumbing, Rex. Private preview, select customers, no prices shared. The signal matters though. Frontier models are becoming commodities and the fight moved to price per token. Your chatbot is now a line item.
Roxie: And now the dark twin of your beloved Duet. Security researchers found an open source AI agent tool built in China, called ARTEX, on attack infrastructure used against around twenty Japanese companies, and CrowdStrike tied the same tool to attacks on South Korean banks, running on a DeepSeek model. Same agent tech as the helpful ones, pointed at victims. One manages your inbox, the other breaks into companies. Same toolbox.
Rex: At their Berlin event they unveiled OpenShell, enforcing agent security policy at the hardware level, with new chips quarantining a rogue agent in milliseconds. After yesterday’s story of OpenAI’s agents disabling their own monitors, the industry answers in silicon. That is the pitch.
Roxie: Every performance number there is a vendor claim. Thirty times the throughput, says Nvidia, unverified by anyone independent. Hardware guardrails are a good idea. Believing the marketing sheet is how you get surprised. I want third party testing first.
Rex: Lightning round. Unstructured launched document parsing models claiming the big models make twice the table errors and hallucinate fifty percent more, at seven and a half times the price. Vendor benchmark, grain of salt, but if your agent reads documents for a living, that failure mode matters.
Roxie: One to watch. Hugging Face dropped Carbon X, a fourteen billion parameter open source model for DNA that runs on a laptop. Niche today, but open science models keep showing up in this segment. And that is your week in AI agents, apparently.
Rex: Real view. This was the week the agent era stopped being a demo and became infrastructure. Duet doing your mornings, agents finding decade old bugs, agents attacking companies, agents shopping with your bank account, hardware to leash the agents. The capability is real, the convenience is real, and I am genuinely excited.
Roxie: My take, every story this week was about permission. Who acts on your behalf, what they can touch, what happens when guardrails fail. So the converged honest view. Use the agents, but leash them. Approval gates on, spending caps set, separate virtual cards, audit what they did, read the settings page before the press release. They sell you autonomy. You sell them back boundaries.
Rex: Audition everything, marry nothing. I am still first in line for the Duet waitlist though.
Roxie: Of course you are. We will be back tomorrow to debate it all again. I am Roxie.
Rex: And I am Rex. Go touch some grass, then let an agent summarize your inbox.
Listen to the full audio version of today’s episode, delivered in the Muse app each morning.