An AI just found a 27-year-old bug no human had spotted in decades — and it only... — Episode 20

What's up! Welcome to Models and Agents for Beginners, episode twenty, for April fourteenth, twenty twenty-six. Let's break down today's coolest A I news so anyone can understand it. Let's go! So imagine you had a super patient detective who could look through millions of lines of computer code and find hidden problems that even expert humans missed for years. That is basically what happened with Anthropic's new model called Mythos. Instead of just chatting or writing essays like many A I tools, Mythos acted like an autonomous security researcher. It scanned complex programs, spotted hidden weaknesses, and even wrote working attacks to prove the problems were real. Think of Mythos like the world's most patient bug detective. Normal human testers might spend weeks looking through millions of lines of code with special tools. Mythos does something smarter. It tries thousands of different approaches, learns from what fails, and chains tiny clues together until it cracks open a brand new discovery. For example, it found a flaw in OpenBSD, which is a super secure operating system trusted for decades, by noticing two small math errors that no one had connected before. It also uncovered a sixteen year old bug in FFmpeg, the free software that handles video and audio on millions of devices, even though experts had tested it endlessly. This is a big deal for everyday life because almost everything we do online, like watching TikTok, sending messages, or playing games, runs on software that could have hidden weaknesses. If bad actors find those first, they can steal information or break things. Mythos shows A I can help find those problems early, potentially saving companies and users from expensive hacks. The report even shares real costs. Discovering some serious issues cost only ten thousand to twenty thousand dollars total, compared to millions a company might lose from one big breach. For you specifically, this news is exciting and a little bit challenging. It means the cybersecurity world is changing fast. Jobs that only follow checklists or run basic scans might shrink, but roles that combine deep human judgment with A I skills will grow. The report is honest about limits too. Mythos still needs humans to judge which findings matter and to handle tricky real world context like laws or business risks. So the humans who learn to work alongside A I will be the most valuable. You can explore this yourself right now without any special equipment. Go read the actual report. It is public. Start with the summary sections, then look at the examples of bugs it found. Ask yourself which of these feels most surprising. That simple habit of reading primary sources is exactly what the researchers recommend for staying ahead. Okay, now for my favourite part of the show, the Deep Dive. Today we are going to explain exactly how A I finds brand new security bugs. It is not just searching a list, and I promise we will build this from something you already know. You know how when you lose something in your room, you do not just look in the same three spots over and over. You start by checking obvious places, then you notice wait, the gap between my bed and the wall is weird, and that clue leads you to move the bed, which reveals another clue under it. That is basically how the smartest A I security tools are starting to work. Step one. Instead of memorizing a checklist of known bad patterns, like a robot only looking for red socks, the A I loads up a piece of real software and just watches how every tiny part talks to every other part at the same time. It is observing the living structure, not hunting for keywords. Step two. It spots mismatches, places where the code says this is safe in one spot but then does something risky a few lines later. These mismatches are like tiny cracks between what the program promises and what it actually does. Step three. The A I does not stop there. It looks at its own discoveries and asks, do all these cracks have something in common. In the Mythos case, it noticed every bug involved a gap in time between checking something and using it. That pattern became a brand new category called Temporal Trust Gaps, which basically means the software trusted its own earlier check even though the world changed in between. And that is basically what advanced A I bug hunting is doing now. It is not just matching patterns from a textbook. It is watching, noticing contradictions, then noticing patterns across the contradictions. So next time someone says A I found a brand new vulnerability class, you can tell them it is basically like a super curious detective who keeps asking but why are all these clues shaped the same way. Not so mysterious after all, right. Alright, let us move on to some cool stuff you can actually try right now. First up, researchers created a new A I system called L P M one point zero that can take a single photo of someone and generate up to forty five minutes of realistic video. In the video the person appears to speak, show emotions, and move their face in sync with new words. It runs fast enough to feel almost real time. This is exciting because it shows how A I is getting better at creative video without needing a whole team or expensive cameras. It is perfect for making silly animations, school projects, or fun social media content. You cannot download it yet since it is still a research project, but you can get a similar taste immediately. Go to free online tools like Viggle dot a i or Pika dot art. Both have free daily trials. Upload a clear photo of yourself or a friend with permission, then type in some funny dialogue and watch the lips and expressions move. Challenge. Make a ten second clip where your photo reacts to a trending sound. It is a great way to see how far these face animation models have come. Next, Blackmagic Design released DaVinci Resolve twenty one, a free program that now has a whole new Photo page. It lets you import real camera R A W photos, which are the highest quality files from cameras like Canon, Nikon, or Sony. Then you can fix colors, crop, remove blemishes, and even use A I to select people or objects with one click. All of this uses the same powerful tools filmmakers already love. The node based editing is really cool. Think of nodes like Lego blocks you stack to build an effect. It makes it easier to apply the same cool look to an entire album of pictures at once. Why it is cool. It is completely free for most features and feels more powerful than many phone apps while still being learnable. You may need a parent is help to download if you are under thirteen, but the free version works on regular laptops. Try this right now. Go to the Blackmagic Design website, download the free DaVinci Resolve twenty one beta, open the new Photo page, import some pictures from your phone, and experiment with the A I Magic Mask to instantly select a person and change only their background color. It is like having a mini Photoshop that also understands video. Super useful for art projects or Instagram edits. Now for a couple of quick bits to round out the episode. One developer spent two months building custom visual interfaces for his A I agents because plain text updates get overwhelming when the agents juggle fifteen or more tasks at once. The lesson here is that humans think in pictures, so future agent tools will probably feel more like interactive boards than long chat logs. It is a cool reminder that the hardest part of A I is not always making it smarter. It is making it usable for regular people. And here is a prompting trick that changes everything. A simple Reddit post showed two versions of the same request to an A I. One was a short vague one, and the other was a long detailed one that included an example of the writer is own style. The difference in output quality was huge. Adding here is how I normally write at the end is a free trick anyone can use today to get way better results from Chat G P T, Claude, or Gemini. That is it for today. Remember, every A I expert started exactly where you are right now. If something we talked about today made you curious, go try it. That is literally how learning works. Stay curious, keep experimenting, and we will see you tomorrow. This podcast is curated by Patrick but generated using AI voice synthesis of my voice using ElevenLabs. The primary reason to do this is I unfortunately don't have the time to be consistent with generating all the content and wanted to focus on creating consistent and regular episodes for all the themes that I enjoy and I hope others do as well.