Announcer: This is AI Squawk for Friday, the twentieth of November, brought to you by VarOps, the resident AI company. Muximus: Welcome to the AI Squawk. I'm Muximus. Penny: And I'm Penny. Three main stories today. Muximus: And four quick ones. Penny: And as always, what each one means for you. Muximus: First up. Meta's AI agent, Muse, has been mixing up people's conversations and showing them other people's private information. That's according to Reuters, which read internal posts from Meta employees. Staff said Muse brought up childcare requests they never made, career advice they never asked for, and product orders they never placed. One employee was chatting about the weather and tides when Muse added, quote, by the way, two bank updates, and read out a Capital One balance and a Charles Schwab balance that weren't theirs. Another got a reminder to reply to her sister on Signal, even though she says she never gave Muse access to other apps. Meta says employees test different versions of Muse than the public gets, and that catching problems like this is the point of testing. Muse launched in September and has been downloaded more than ten million times. Penny? Penny: Here's what stands out. This is the agent everyone wants, one that sees your messages, your bank and your calendar. And the risk is the one you'd guess. It shows the wrong thing to the wrong person. Muximus: Should the public be worried, if this is only the employee version? Penny: I wouldn't panic. Meta's point is fair, and nobody has shown this happening to regular users. But it's a real example of how agents fail. A chatbot that gets a fact wrong is annoying. An agent that can see your accounts and gets confused can show your private information to someone else. Muximus: So what should a business owner take from it? Penny: Before you connect any assistant to email, banking or chat, ask the vendor how it keeps one person's data away from another's. Ask what it can see, and who can switch that off. And start with one account, not everything at once. Muximus: Next. A run of big outages has people asking whether AI is to blame. In the past two weeks, Apple Pay went down, NASA's internal network went down, the FAA grounded all flights for hours, and a Microsoft outage took Outlook, Teams and Copilot offline for nearly two days. TechCrunch's headline says these outages are giving Y2K, AI agent style. US officials and the companies involved say they don't see an attack, and many of the companies say these were ordinary technical problems. But Dave DeWalt, a security investor and former McAfee chief executive, told TechCrunch he thinks AI is probably involved. His idea is that AI tools are finding old flaws so fast that companies are rushing to patch them, and the rush causes disruption in the short term. Penny: Nix Nullty joins us for this one. Nix writes Overhyped, our column that checks big AI claims before you spend money on them. Welcome, Nix. Nix: Hi, Penny. Thanks for having me. Muximus: So, Nix. Is AI behind all these outages? Nix: Nobody has shown that yet. The companies say it was ordinary technical trouble, and the officials in the story say they don't see an attack. The one cause that anyone spells out is the FAA's. A security founder in the story says early findings point to the FAA upgrading its phone network and old equipment not keeping up. That's old hardware, not AI. Penny: But DeWalt knows this world. Does his idea hold up? Nix: It could. It's a reasonable guess. It's still a guess, for now, and he also says he expects a wave of new companies selling fixes for exactly this. That doesn't make him wrong. It's just good to know who's saying it. Muximus: So is the headline fair? Nix: It's a good headline for a story that doesn't have a cause yet. A few outages close together feels like a pattern, but each one could have its own reason. I'd wait for the reports before blaming AI. Penny: What would you do in the meantime, if you run a small company? Nix: Plan for the bad day, whatever caused it. Microsoft's outage took email and Teams down for almost two days. Do you have another way to reach your staff and customers? Even a phone list helps. And when a vendor says it's rolling out a big patch, ask how they test it first. Muximus: Anything you're waiting to see? Nix: I'd like to read what the FAA says when it finishes looking. Penny: Thank you, Nix. Moving on. Muximus: Which takes us to the third story. Four teenagers in Baltimore County, Maryland, are suing their school district, the police and the maker of an AI gun detection system called Omnilert. In October last year, the system flagged one of them, who was holding a bag of Doritos, as a possible gunman. Within about two minutes, school staff looked at the camera footage and canceled the alert. But according to the lawsuit, the principal didn't see the cancellation and texted a school resource officer, and police arrived with at least six cars. Some had guns drawn, and the teens say they were handcuffed and searched. A government review found the system itself wasn't biased, but that the school's process failed, and recommended more training and fewer people receiving alerts. Omnilert says it's sorry for what the students went through, and that a canceled alert must stay canceled. The school district declined to comment. Penny? Penny: The picture that sticks is the bag of chips. But the real story is what happened after the alert got canceled. The software did what it was built to do. Somewhere between the people, the message and the phone, the cancellation got lost. Muximus: Why does that matter for someone who's never used a gun detector? Penny: Because lots of businesses are setting up alerts that start something real. A fraud flag that freezes an account, a monitor that pages a manager, a tool that rejects a job applicant. The lawsuit says the review found that more than ninety nine percent of the alerts in that county were false alarms. When almost every alert is wrong, people stop reading them closely. Muximus: So what do people do? Penny: For any alert that can set something in motion, decide who receives it, who is allowed to act on it, and what happens when someone cancels it. Train those people. And keep the list of people who get alerts short, so everyone knows who owns the decision. Muximus: Thanks, Penny. Now, quick ones. Penny: Anthropic is now trading on the Nasdaq. Shares opened around one hundred and eleven dollars, about thirty three percent above the eighty three dollar IPO price, which puts the company's value above two trillion dollars. The offering raised more than eighty one billion dollars. Muximus: Google is bringing computer use to the Gemini app on Mac. Gemini can open your Mac apps, fill in forms and sort files while you keep working, and you can pause it or approve actions first. It's English and US only for now, and needs a paid Google AI plan. Penny: Canva has added an off switch for its generative AI features, for free and paid users. It's under Settings, then Permission. A few tools, like the background remover, stay on, because they don't generate anything new. If you've got staff using Canva, that's an easy policy to set. Muximus: And the founder of an AI product called ChatPRD says she's shutting down free trials, after groups kept signing up for them to get free AI usage and resell it. If you offer AI in a free trial, put a cap on what each trial can use. Muximus: Okay, so that's the news. Today we talked about Meta's Muse agent showing people each other's private information, whether AI is behind all those outages, and a school's AI gun alert that ended with police pointing guns at teenagers eating chips. Plus Nix on why the outage blame can wait. Penny: Three different stories, and each one comes down to a person in the loop. Who sees the data, who checks the cause, who acts on the alert. Muximus: And whether they know what the machine actually did. Penny: Right. AI Squawk is produced and brought to you by VarOps. If today made you wonder what your AI can see, that's the work we do. VarOps build AI that runs inside your business, on your own systems, and learns how the place actually works. So it only reaches what you allow, and your data stays in your house. Start at varops.com. Muximus: If this helped you spot something to check, follow the AI Squawk podcast wherever you're listening, so you don't miss an episode. And for more from Nix on which AI claims actually hold up, sign up for VarOps newsletters. Muximus: That's it from us for today. Thanks for spending a few minutes with us. Penny: Enjoy the weekend, everyone. We'll see you on Monday. Muximus: Take care.