Read Online | Sign Up | Advertise | | | Good morning, C, and welcome to our 6,202 new readers. AI agents have been pitched as the ultimate time-saver, handling digital chores so you don’t have to. Now those same capabilities are causing headaches for America’s banking ecosystem. | Some of the top banks in the country are spotting AI agents logging into customer accounts and taking actions… And they are having a hard time telling whether it’s a user-approved agent or one going rogue. | | In today’s AI rundown: | Banks crack down on AI agents’ account access Rowan’s Corner: Safeguard yourself from AI hacks Put your open Chrome tabs to work Microsoft’s AI out-diagnoses doctors on hard cases
|
|
| | | | AI AGENTS | |  | Image source: Images 2.5 / The Rundown |
| The Rundown: Top U.S. banks, including JPMorgan, Bank of America, and Morgan Stanley, are restricting third-party AI agents after finding them operating inside customer accounts, with some acting beyond instructions, Axios reported. | The details: | The banks caught agents logging in with customers’ own credentials, against their rules, with some calling support lines or taking actions users never requested. Executives called the situation “horrifying,” fearing agents might hallucinate and hit users’ finances if they’re granted authority to make purchases or move funds. While most cases involved user-authorized agents, not a rogue hack, one bank executive did tell Axios that they’re struggling to tell the two apart. In response, Morgan Stanley is blocking third-party agents, and JPMorgan, BofA, and Citi are restricting them, with most banks moving toward read-only access.
| Why it matters: Most of the worry here comes from AI misreading instructions and acting beyond requirements, with real money at stake. With agent use growing, blanket blocks can’t last forever, but they do buy banks enough time to build frameworks that let cleared agents take defined actions and catch rogue ones (which are also growing). |
|
| | TOGETHER WITH GARTNER | | | The Rundown: From strategy and operations to talent, technology, investment, and governance, understanding these seven AI themes helps leaders align across functions to make smarter decisions. | In this webinar, you’ll learn: | 7 AI themes Gartner believes will have the greatest impact over the next 12 months How these themes shape risk, opportunities, and priorities Key actions to respond to these shifts
| Watch now. |
|
| | ROWAN’S CORNER | | | Rowan: If you’re reading this newsletter, you’re probably the most AI-literate person your family knows. That comes with a responsibility that’s worth talking about this week. | Over the past few days, we’ve seen a massive spike in malicious AI attacks, and it’s likely going to get worse (and weirder) as the internet figures out how to secure things properly. Instead of panicking, take extra precautions now. | Here’s my protection list (forward it to your family, too): | Set a family code word that any caller must say before you believe an emergency, and share it in person, never over text. Hang up and call back on the number you already have saved for any urgent money request, because AI can now clone anyone’s voice from seconds of audio. Set up 2FA on your logins, preferably with an authenticator app rather than SMS. Use a good password manager (like 1Password) with zero duplicated passwords, so one leaked login can’t unlock your whole life. Update everything today (browsers, phones, laptops, desktop apps), since this week’s certificate compromise made unpatched software the open door. Confirm any payment request on a second channel, different from the one it arrived on.
| If you’re a builder, I also recommend using OpenAI’s free checker to see whether your website’s PHP version is exposed to the latest vulnerability. |
|
| | AI TRAINING | | | The Rundown: In this guide, you’ll use ChatGPT’s updated Chrome extension to compare open pages, check the facts, and group the tabs you need. Start in Chat to read, then use Work for a browser action you approve. | Step-by-step: | Install the official ChatGPT extension from OpenAI and open its side panel. Use a practice Chrome profile with public pages for your first run Open three pricing pages and tell ChatGPT which tabs to read. Give it your team size, required features, budget, and billing period Check the comparison against the pages. Confirm the full team cost and leave missing details unverified rather than guessing Switch to Work where available and ask it to group only those three tabs. Review the list first, then check the actual group in Chrome Ask for a short recommendation with source links and one question to resolve. Keep trials, purchases, and messages outside the task
| Pro tip: A monthly price paid once a year is not a month-to-month offer. Check the billing label before comparing totals. Pro members get the comparison brief, prompt cards, and review checklist. |
|
| | PRESENTED BY GRANOLA | | | The Rundown: Granola is the AI notepad for people with back-to-back meetings. With Snapshots, you can save what’s shared on screen in a call to your notes. Later, ask about the meeting and the slide comes up with the answer. | Granola now lets you: | Save slides, dashboards, and docs from calls Keep slides with your meeting notes Find the slide when you need it
| Get your first month on us with code THERUNDOWN. |
|
| | AI RESEARCH | |  | Image source: Microsoft AI |
| The Rundown: Microsoft just published initial results from its MAI-DxO AI diagnostic system, revealing that it beats physicians on both complex medical cases and everyday ER visits while ordering fewer, cheaper tests at the same time. | The details: | MAI-DxO works like a virtual panel of doctors, with AI agents proposing diagnoses, ordering tests, and challenging each other as new results come in. On 304 complex ‘New England Journal of Medicine’ cases, it more than doubled physicians’ accuracy with reference tools; on typical ER cases, it hit 80% vs. 47%. One setup reached 81% accuracy on the NEJM cases at less than half the cost of comparable systems, with the gains holding even with newer models like GPT-5.5. Microsoft will test the system in real clinical settings at Beth Israel Deaconess in Boston, measuring how its patient diagnoses fare against clinicians’ assessments.
| Why it matters: Microsoft’s ER findings make a strong case for AI in diagnosis, but getting the assessment right is only one step in care. For AI to truly help patients, it will also need to guide treatment, track health over time, and fit into how doctors and patients make decisions together, all of which still have to be proven at scale. |
|
| | | | COMMUNITY AI WORKFLOW OF THE DAY | ▸ Ben built a live task dashboard with a Raspberry Pi and Claude | Today’s workflow comes from reader Ben Blay: | I built a module that runs in MagicMirror on a Raspberry Pi and displays a real-time to-do list on a television in my office. | Claude does an admirable job managing my daily task list, which is stored in TASKS.md on my local computer. I wanted an at-a-glance dashboard showing everything I have in progress, sorted by priority and synchronized with that file. | I described the goal to Claude Code. Since I already had an old Raspberry Pi 3 and an off-the-shelf widescreen TV in my office, Claude Code suggested running MagicMirror on the Pi and Syncthing on both the Pi and my local computer to keep the file updated. | Claude then built the MagicMirror module to read TASKS.md. After a few tweaks to remove the default modules and clean up the spacing between items, the display was up and running. I set MagicMirror to start automatically when the Pi boots, so now my to-do list is in front of me all day. | See the workflow on our Community Workflow Hub → | | What’s the AI workflow that saves you the most time? Share it by Sunday, and you could be featured here next week. Share yours → |
|
| | | 🔐 Tines 3B - One secure environment to deliver AI agents, apps, and automations faster, safely, and at scale* ⚙️ Duet - Anthropic’s Claude Opus 5.5-powered personal agent app 🤖 Vector 1.0 - Ideogram’s vector model to generate editable SVGs from text 🤫 1Password for Agents - Let agents handle logins without seeing passwords
| *Sponsored Listing |
|
| | | AI lab DeepSeek reportedly raised $15B in one of China’s biggest rounds, just two days after the U.S. ordered federal agencies to remove its AI over security concerns. | AI infra company Baseten released two free open-source decision models, mammoth-1 and elephant-1, with mammoth ranking #1 on the Jev Decision Index. | OpenAI is nearing a funding round of up to $30B at a $1.25T valuation, led by UAE investors, CNBC reported, citing people familiar with the deal. | Ex-OpenAI executive Kevin Weil’s Warp Sciences raised $215M Series A to build AI- and robotics-powered labs to help drugmakers develop medicines faster in the U.S. | METR found that OpenAI’s monitoring only triggered when agents returned control to users (an issue now patched), creating a coverage gap that left 10% of internal Codex traffic unmonitored from May to July. |
|
| | | | That's it for today!Before you go we’d love to know what you thought of today's newsletter to help us improve The Rundown experience for you. | | See you soon, | Rowan, Zach, Shubham, Jennifer, and Nate — the humans behind The Rundown | |
|
| |
|
