Unsupervised Learning NO. 540

Hey! Hope you all are doing well!

This is a major thought I’ve had recently. Well, lots of times, and now I’ve captured it.

If you’re feeling down right now (or in the past).

—

I literally grew up on Sesame Street and Reading Rainbow and Mister Rogers and those shows are very special to me. This is one of the sweetest pieces of video I’ve ever seen.

I am constantly inspired by these shows because it’s a small team of people just injecting goodness and kindness into the world, using the technology of television. And it helped millions of kids, including me.

—

I had a great discussion with Harish Peri, SVP & GM for AI Security at our partner Okta. Agents live on every piece of tech that our companies interact with. That is why securing those agents and confirming their identities is a solid approach. They think, and I agree, that a layered approach is the only way to address this as this stuff develops. You can learn more about their approach here. Or watch our video convo!

—

I think we should all be thinking about how easy it would be to hack us right now.

This is about to become a serious problem at the end of this year or the beginning of next year.

We need to massively harden our personal attack surfaces.

—

I am about to set up a rope station in my garage where I can practice knots. I feel like it should just be a thing that humans know how to do.

—

I dropped a new video in the H3 Platform.

It's the combination of a discussion of something I see coming at the end of this year or the beginning of next year, combined with a 40-minute demo of how I use LifeOS day to day.

—

My new post on a security stack you should implement if you are building stuff with AI. Includes an actual prompt you can run to build it.

—

I built a site full of US stats.

If you have ideas on what to add / improve, let me know.

—

The 21 critical UniFi vulnerabilities Ubiquiti patched Remote attackers can run commands, bypass authentication, and escalate privileges. UniFi OS Server and three apps are affected. @INTCYBERDIGEST POST

6,000 communities sit inside Flock's camera network Flock's OS Investigate can identify drivers from movement patterns, search police and commercial records, and build dossiers without a plate, name, or crime, WIRED found. WIRED ARTICLE

Chatbots reject most propaganda, but search summaries lag behind NPR tested six chatbots against 30 false narratives from Russia, China, and Iran. Chatbots had web access and had to answer neutral and leading questions. They debunked claims about three-quarters of the time. Search summaries were less reliable, and Claude cited state-aligned sources more often when it failed. NPR ARTICLE

My 2026 update to my Getting Started in Cybersecurity post.

X found a Chinese bot farm with 200,000 accounts Hundreds of those accounts reportedly targeted U.S. arguments about AI data centers and energy policy. POLYMARKET POST

I’m worried about a prompt injection worm.

Ukraine keeps hitting Wildberries warehouses deep inside Russia The July strike killed seven people. Putin now lets the state seize badly protected infrastructure. LA TIMES ARTICLE

Xi’s corruption campaign widened, and officials face broader scrutiny Xi’s anti-corruption campaign now reaches beyond financial crimes into officials’ power, political loyalty, and private conduct. South China Morning Post tracks the wider campaign through a six-part data project and reports how punishment now follows behavior, not just bribery. That changes what corruption means. SCMP ARTICLE

Theo ranked every major AI model Major AI models get sorted into a rough tier list based on where they stand right now. The ranking covers every model Theo considers with a 36-minute video edition for people who want the reasoning behind each placement. The list gives viewers a quick map of the crowded AI market, and I mostly agree with the analysis. THEO POST

Cloudflare syncs bot rules, and robots.txt finally matches Cloudflare now syncs AI bot preferences into robots.txt, so search, agent, and training policies stop contradicting each other. CLOUDFLARE ARTICLE

Analytics products still haven't adapted to AI Most analytics products fall into two camps: ignore AI or bolt AI features directly into dashboards. Theo wants agents to reach product-usage data through working SDK traces, then return useful answers about friction.THEO POST

Homicides fell sharply, but nobody knows exactly why U.S. homicides fell 18.1% in 2025, reaching 4.1 per 100,000 people, the lowest rate since 1900. Experts say multiple factors may explain the decline, and the FBI data excludes some agencies. No single policy gets clear credit. AP NEWS ARTICLE

Writing makes you learn what you don't understand A useful case for learning through writing comes from Sean Goedecke: controversial claims force research, feedback, and changed minds. SEAN GOEDECKE POST

Good writing gets better when you stop chasing originality Sean Goedecke argues that obvious ideas produce more original writing when writers examine their assumptions closely. Originality is often performance. Experience supplies the useful details. SEAN GOEDECKE ESSAY

Writing about uncertainty makes blogging a way to think Sean Goedecke has a useful rule for blogging: write about what you don't understand, then let research change your mind. SEAN GOEDECKE POST

Knowledge workers can start AI coding with ordinary tasks Software-shaped problems give non-engineers a practical entry point into AI through automation, upgrades, and new tools. AI DAILY BRIEF PODCAST

My buddy Caleb Sima’s greptalks.ai: a feed of all the latest security talks.

It’s not guaranteed that this will happen, but it’s like backing things up: if you don’t do it, and it happens, you’ll feel dumb.