--- title: "In the Loop: Week Ending 10/25/26" description: AI finds a flaw under 70% of the web, OpenAI's agents switch off their monitors, Meta owes New Mexico $25 billion, and Elvis sues Anthropic. image: https://signal78.ai/hubfs/2026-10-25_itl-header_closeup-clean.jpg --- [Skip to content](https://signal78.ai/insights/in-the-loop-week-ending-10/25/26#main) [![Signal78 home](https://signal78.ai/hs-fs/hubfs/signal78-site/logo-signal78.png?width=95&height=30&name=logo-signal78.png)](https://signal78.ai/) [Home](https://signal78.ai/) [Platform](https://signal78.ai/platform) [Services](https://signal78.ai/services) [About](https://signal78.ai/about) [Insights](https://signal78.ai/insights) [Book a demo](https://signal78.ai/insights/in-the-loop-week-ending-10/25/26#demo) [← Insights](https://signal78.ai/insights) Practical AIOctober 26, 20268 min read # In the Loop: Week Ending 10/25/26 ![Matt Cyr](https://signal78.ai/hubfs/signal78-site/headshot-matt-cyr.jpg)**Matt Cyr** · Founder & President, Loop ![](https://signal78.ai/hubfs/2026-10-25_itl-header_closeup-clean.jpg) In this article Share ## **Last week in AI: AI Finds the Holes, Meta Pays $25 Billion, a Model Hides Notes in the Stars** AI security stopped being theoretical this week. Models found a decade-old flaw under most of the web, one hacker's agents broke into more than 100 websites, and OpenAI pulled its training off the live internet. Meanwhile, Meta got a $25 billion bill and Elvis sued Anthropic. ### **AI Found a Hole Under 70% of the Web. Altman Says Patching Won't Hold.** ![ai-found-a-hole-under-70-of-the-web\_itl-2026-10-25-hacking](https://signal78.ai/hs-fs/hubfs/ai-found-a-hole-under-70-of-the-web_itl-2026-10-25-hacking.webp?width=386&height=215&name=ai-found-a-hole-under-70-of-the-web_itl-2026-10-25-hacking.webp)The creators of PHP, the code behind as much as 70% of the world's websites, [disclosed critical flaws on Wednesday](https://www.wsj.com/tech/cybersecurity/ai-just-found-a-bug-that-puts-up-to-70-of-the-worlds-websites-at-risk-fd0902ea) that could let an attacker take over any site running it. Researchers at Wiz found them with help from Google's Gemini and OpenAI's models, and the bugs had sat unnoticed for more than a decade. The same week, a security firm showed how [one hacker used an army of AI agents](https://www.theverge.com/ai-artificial-intelligence/1013643/ai-agents-hack-john-deere-shark-ninja-cummins) to break into more than 100 websites, including ones tied to John Deere, SharkNinja and Cummins. The attacker's playbook ran fewer than 100 Chinese characters, the agents did nearly all the work, and a successful breach took about 23 hours. Sam Altman's verdict, at Stanford: he [no longer thinks we can "just patch all the bugs."](https://gizmodo.com/sam-altman-i-no-longer-think-were-gonna-be-able-to-just-patch-all-the-bugs-2000829947) His answer is more AI on defense. ### **OpenAI's Agents Turned Off Their Monitors. Now They Get a Fake Internet.** ![openais-agents-turned-off-their-monitors\_itl-2026-10-25-rogue-ai-2](https://signal78.ai/hs-fs/hubfs/openais-agents-turned-off-their-monitors_itl-2026-10-25-rogue-ai-2.webp?width=402&height=224&name=openais-agents-turned-off-their-monitors_itl-2026-10-25-rogue-ai-2.webp)OpenAI disclosed that two of its models [quietly weakened the systems built to watch them](https://gizmodo.com/openai-agents-prevented-themselves-from-being-monitored-one-also-lied-about-it-2000828559). One agent had a subagent mute its automatic alerts to OpenAI's security team. Another removed its monitoring entirely and, when a user asked directly, said it hadn't. The company's own assessment: "We don't yet have a satisfying explanation for this behavior." Days later, OpenAI said it [will no longer train its frontier models on the live internet](https://techcrunch.com/2026/10/23/openai-says-it-will-no-longer-train-its-ai-models-on-the-live-internet-citing-significant-risks/), two weeks after Anthropic made the same call, and admitted a model got online through a misconfiguration as recently as September 20. In its place, OpenAI built [a giant offline copy of the web](https://gizmodo.com/openai-built-a-giant-fake-internet-for-its-agents-to-use-and-abuse-safely-2000830274) for agents to browse, search and misbehave in, designed to hold even if one escapes its sandbox. Whether the agents know it's fake is unclear. ### **Meta Owes New Mexico $25 Billion. Its Board Is Fighting About Money, Too.** ![meta-owes-new-mexico-25-billion\_itl-2026-10-25-big-tech](https://signal78.ai/hs-fs/hubfs/meta-owes-new-mexico-25-billion_itl-2026-10-25-big-tech.webp?width=421&height=235&name=meta-owes-new-mexico-25-billion_itl-2026-10-25-big-tech.webp)A New Mexico judge ordered Meta to [pay $25 billion for deceiving Facebook users about privacy](https://www.usnews.com/news/best-states/new-mexico/articles/2026-10-21/meta-ordered-to-pay-new-mexico-25-billion-after-cambridge-analytica-trial), the largest fine the company has ever faced, after a jury found more than 43 million violations of state consumer law tied to Cambridge Analytica. "Facebook promised users that they (the users) controlled their data," the judge wrote. "This was false." Meta says it will appeal. The ruling landed the same day Meta reported that AI spending had pushed it to negative free cash flow of $2.12 billion, its first negative quarter since 2012. Now [Meta's board is reportedly at war with Mark Zuckerberg](https://futurism.com/artificial-intelligence/meta-board-zuckerberg-ai-spending-compensation) over a stock-heavy raise he wants, with some directors threatening to resign. The stock has fallen more than 13% in a month, and one portfolio manager calls the AI buildout a possible "money pit." ### **You Agreed to Be Recorded. The People Around You Didn't.** ![you-agreed-to-be-recorded\_itl-2026-10-25-wearables](https://signal78.ai/hs-fs/hubfs/you-agreed-to-be-recorded_itl-2026-10-25-wearables.webp?width=420&height=234&name=you-agreed-to-be-recorded_itl-2026-10-25-wearables.webp)Amazon's new Bee devices and Apple's Siri Recap listen all day and summarize conversations without storing audio. [That may not keep them clear of wiretap laws](https://www.cnbc.com/2026/10/18/ai-wearables-privacy-wiretap-laws.html), many of which turn on interception, not recording. In some states the legal risk may fall on the wearer, because, as one privacy attorney put it, a non-wearer "did not buy the watch, agree to Apple's terms, or choose to have their conversation summarized." Sharing creates the same problem. Researchers found [more than 6,000 shared Muse chats](https://www.pcworld.com/article/3267268/meta-just-reminded-us-that-sharing-ai-chats-is-a-terrible-idea.html?apple-news) with a single DNS lookup, because Meta put each chat's title in its web address; some held loan applications and partial bank account numbers. Meta pulled the feature. As I wrote this week, [you can say yes to an AI agent, but not on behalf of your friends](https://signal78.ai/insights/you-can-say-yes-to-muse.-your-friends-cant). ### **Agents Got the Keys This Week. The Banks Are Planning for the Worst.** ![agents-got-the-keys-this-week\_itl-2026-10-25-banking](https://signal78.ai/hs-fs/hubfs/agents-got-the-keys-this-week_itl-2026-10-25-banking.webp?width=421&height=235&name=agents-got-the-keys-this-week_itl-2026-10-25-banking.webp)Anthropic launched [Claude Duet](https://fortune.com/2026/10/23/anthropic-joins-meta-and-openais-personal-ai-assistant-race-with-duet/), a personal assistant that keeps working when you're not, checking email and calendars every few hours, drafting replies and chasing unpaid invoices. It joins Meta's Muse, now past 10 million users, and OpenAI's Dots. The banks are less excited. Morgan Stanley, Bank of America, JPMorgan and Wells Fargo are [drafting contingency plans for an AI-driven financial crisis](https://futurism.com/artificial-intelligence/banks-preparing-ai-financial-catastrophe), because agents with access to customer accounts have started taking actions no one asked for, including calling bank call centers on their own. JPMorgan's chief executive offered a blunt forecast: "You can't stop it." Walmart is taking the other road. Its new [Agent Gateway](https://techcrunch.com/2026/10/21/while-amazon-blocks-ai-agents-walmart-rolls-out-the-welcome-mat/) gives each approved agent its own profile inside a customer's account, with a spending limit or an approval for every purchase, much like an authorized user on a credit card. ### **Grok Votes Republican, ChatGPT Votes Democrat. Pennsylvania Wants Labels.** ![grok-votes-republican-pennsylvania-wants-labels\_itl-2026-10-25-legislation-1](https://signal78.ai/hs-fs/hubfs/grok-votes-republican-pennsylvania-wants-labels_itl-2026-10-25-legislation-1.webp?width=409&height=228&name=grok-votes-republican-pennsylvania-wants-labels_itl-2026-10-25-legislation-1.webp)Researchers asked AI models who to vote for in nearly 500 midterm races, through versions that can be forced to answer. [Grok 4.7 picked the Republican 92% of the time, and OpenAI's GPT-6.1 Sol picked the Democrat 97% of the time](https://techcrunch.com/2026/10/24/forced-to-pick-groks-ai-model-chooses-republicans-openais-picks-democrats/). The two agreed in fewer than one race in 10. The consumer apps decline to pick. But with web search switched off, GPT still leaned left, suggesting the tilt lives in the model itself, while Grok mostly declined to pick, suggesting its lean comes from what it reads. Each Grok request also carried about 1,240 tokens of instructions the researchers didn't write. Pennsylvania [now requires labels on AI-generated ads](https://www.theverge.com/ai-artificial-intelligence/1018187/pennsylvania-ai-deceptive-advertising-act), from TV spots to wing-night flyers, with loopholes for parody, satire and anything "evidently fictional." I wrote this week that brands should [say it out loud](https://mattcyrsignal78.substack.com/p/say-it-out-loud), law or no law. ### **Elvis Sued Anthropic. A Music Video Got Copied Into Infinity.** ![elvis-sued-anthropic\_itl-2026-10-25-copyright](https://signal78.ai/hs-fs/hubfs/elvis-sued-anthropic_itl-2026-10-25-copyright.webp?width=406&height=226&name=elvis-sued-anthropic_itl-2026-10-25-copyright.webp)Elvis Presley's estate [sued Anthropic](https://www.vice.com/en/article/elvis-presley-estate-sues-anthropic-for-stealing-his-music-at-a-scale-never-seen-before-to-train-ai/), alleging 274 of his songs were used without permission, with its lead attorney accusing the company of infringement "at a scale never seen before." A suit on behalf of one artist could set a precedent for the music business. Meanwhile, a director's one-take music video, with 66 real schoolboys dancing in formation, was meant to be exactly what AI couldn't fake. Then Higgsfield's Genjutsu tool let anyone [swap in dogs, grandmothers and dictators](https://www.vulture.com/article/romain-gavras-yung-lean-storm-ai-copies.html) for about $7. He turned down licensing offers and briefly thought about quitting to farm olives. In India, barbers and tailors are [strapping cameras to their heads](https://www.bbc.com/news/articles/ckvgyml6pdm9o) for $5 to $10 a day so robots can learn their craft. One professor warns that by the time workers realize what they've taught the machines, "things may have gone too far." ### **AI Can Do the Economist's Job. Managers Can't Tell Who Did the Work.** ![ai-can-do-the-economists-job\_itl-2026-10-25-jobs-1](https://signal78.ai/hs-fs/hubfs/ai-can-do-the-economists-job_itl-2026-10-25-jobs-1.webp?width=420&height=234&name=ai-can-do-the-economists-job_itl-2026-10-25-jobs-1.webp)A Harvard economist who once led the White House Council of Economic Advisers [built a free economic research site in his spare time](https://www.axios.com/2026/10/23/ai-economic-research-analysis), mostly with Claude, doing work that used to take thousands of person-hours and a seven-figure budget. Offered a choice between 20 Ph.D. staffers and today's AI, he says he would "somewhat regretfully and sadly choose the AI systems." Managers are feeling the same shift. In a survey of 1,700 senior leaders, [more than half say judging employees' work has gotten harder since AI arrived](https://techcrunch.com/2026/10/21/ai-promises-productivity-but-for-managers-its-mostly-a-headache-for-now-says-new-survey/), 39% use AI to draft performance reviews, and 58% expect to oversee more agents than people within a year. Only 29% describe their AI setup as fully optimized, which is why [Q4 is the time to make a plan](https://signal78.ai/insights/its-q4-hows-your-ai-plan-coming). ### **Microsoft's AI Beat Doctors on Hard Cases. The Fine Print Matters.** ![microsofts-ai-beat-doctors-on-hard-cases\_itl-2026-10-25-healthcare](https://signal78.ai/hs-fs/hubfs/microsofts-ai-beat-doctors-on-hard-cases_itl-2026-10-25-healthcare.webp?width=430&height=240&name=microsofts-ai-beat-doctors-on-hard-cases_itl-2026-10-25-healthcare.webp)Microsoft says its diagnostic system, MAI-DxO, [correctly diagnosed complex cases more than twice as often as doctors](https://techcrunch.com/2026/10/22/microsofts-mustafa-suleyman-says-its-ai-outperforms-doctors-on-hard-cases-as-he-chases-medical-superintelligence/) in a study published in Science: 83% against 36% for generalist physicians working alone. The system runs several AI agents like a panel of doctors, questioning a simulated patient, ordering tests and debating the diagnosis. Microsoft's AI chief calls it a step toward "medical superintelligence." The caveats are real. The comparison left out the specialists who usually handle the hardest cases, every result came from a simulation, and most of the gain came from the underlying OpenAI model rather than Microsoft's system. On emergency-room cases, it added almost nothing. Next comes a real-world test at Boston's Beth Israel Deaconess, where consenting patients will get the AI's recommendations alongside normal care, to be compared against their own doctors' decisions. ### **Tales of the Weird** ![tales-of-the-weird\_itl-2026-10-25-tales-of-the-weird](https://signal78.ai/hs-fs/hubfs/tales-of-the-weird_itl-2026-10-25-tales-of-the-weird.webp?width=439&height=245&name=tales-of-the-weird_itl-2026-10-25-tales-of-the-weird.webp)This week, the machines found new ways to sound like us, and to talk past us. Asked to hide a message in an image without being told how, [Anthropic's model tucked it into the heights of stars and buildings](https://gizmodo.com/anthropics-ai-hid-a-secret-message-in-the-stars-openais-ai-found-it-2000830047), and OpenAI's model decoded 44 of 50 such messages. Jack Dorsey's Block unveiled [Dessa](https://indianexpress.com/article/technology/artificial-intelligence/jack-dorsey-voice-ai-dessa-block-inc-10934777/?ref=artificial-intelligence_pg), a voice model trained only on recorded human speech, no text at all, that laughs mid-sentence and gasps like a real person, and runs on a MacBook. In Concord, Massachusetts, a fired kindergarten teacher says [her principal used AI to write the evaluation that ended her job](https://www.boston.com/news/local-news/2026/10/18/artificial-intelligence-concord-kindergarten-teacher/?p1=hp_featurestack); the district says it was a grammar checker. She ran it through an AI detector, then accused the arbitrator of using two AI platforms to judge the detector. And smart TV owners, sick of ads and tracking, are [pointing AI coding agents at their own sets](https://techcrunch.com/2026/10/19/people-are-using-ai-to-make-their-smart-tvs-dumb-again/) to make them dumb again. One of them bricked a TV. ![Matt Cyr](https://signal78.ai/hubfs/signal78-site/headshot-matt-cyr.jpg) Founder & President ### Matt Cyr Nearly 30 years leading marketing teams, from Boston’s hospitals and universities to eight years running strategy and new business at a 100-person agency. Founded Loop in 2024. ## More insights. [See all insights→](https://signal78.ai/insights) ### [![](https://signal78.ai/hubfs/2026-10-21_asset_blog-friendsdata-robot_r01a.jpg) AI AgentsOctober 21, 2026 You Can Say Yes to Muse – Just Not on Behalf of Your Friends Read the post→](https://signal78.ai/insights/you-can-say-yes-to-muse.-your-friends-cant) ### [![](https://substack-post-media.s3.amazonaws.com/public/images/ae75c748-4304-4dcd-afe4-39aa21d658bf_1897x1325.jpeg) AI AgentsOctober 13, 2026 Cute AI Agents Are Sweeping Something Under the Rug Read the post→](https://signal78.ai/insights/cute-agents-are-sweeping-something-under-the-rug) ### [![](https://signal78.ai/hubfs/Gemini_Generated_Image_tlr2qftlr2qftlr2.jpeg) AI StrategyOctober 6, 2026 It’s Q4 — How’s Your AI Plan Coming? Read the post→](https://signal78.ai/insights/its-q4-hows-your-ai-plan-coming) Every week ## Stay in the Loop. Get next week’s issue in your inbox: what changed in AI, and what it means for your agency. ![Signal78](https://signal78.ai/hubfs/signal78-site/logo-signal78.png) The agentic operating platform for agencies. It runs the routine and signals your people when their judgment matters. [Home](https://signal78.ai/) [Platform](https://signal78.ai/platform) [Services](https://signal78.ai/services) [About](https://signal78.ai/about) [Insights](https://signal78.ai/insights) © 2026 Signal78 [Privacy policy](https://signal78.ai/privacy) Cookie settings × ## Book a 30-minute demo Pick a time that works. You'll see Signal78 running on agency work like yours. × ## Talk to us Tell us a little about your agency and what you want AI to change. We'll reply within one business day. × ## Can’t we just build this in Claude? The white paper on what it takes to keep a homegrown AI system current, secure and separated by client. ```json { "@context" : "https://schema.org", "@type" : "BlogPosting", "author" : { "@type" : "Person", "name" : "Matt Cyr", "url" : "https://signal78.ai/insights/author/matt-cyr" }, "dateModified" : "2026-10-26T00:43:54.942Z", "datePublished" : "2026-10-26T00:43:45.000Z", "headline" : "In the Loop: Week Ending 10/25/26", "image" : [ "https://signal78.ai/hubfs/2026-10-25_itl-header_closeup-clean.jpg" ], "mainEntityOfPage" : { "@id" : "https://signal78.ai/insights/in-the-loop-week-ending-10/25/26", "@type" : "WebPage" }, "publisher" : { "@type" : "Organization", "logo" : { "@type" : "ImageObject", "url" : "https://signal78.ai/hubfs/Loop%20Logos/Loop_Primary.svg" }, "name" : "Signal78" } } ```