Now With More Post Mythos Apocalypse!  🔗 **Source:** [Now With More Post Mythos Apocalypse!](https://share.snipd.com/episode/4a63ba3c-029d-41a0-aa7e-f361daa55838) ## Highlights - **Looping LLMs Enable Long Horizon Autonomous Tasks** - Looping LLMs autonomously repeat actions until success criteria are met, enabling very long-horizon tasks like persistent penetration attempts. - Christopher S. Penn notes Opus/GLM models can loop hours to days, turning a high-level goal into continuous automated efforts. Transcript: Christopher S. Penn So Mythos is apparently a model that is somewhat better in terms of intelligence than Opus 4.6. It's probably the equivalent of Opus 5. However, it is using a technology called looping LLMs. And what this is, is these are tools that are not only good at tool handling, but also recognizing that they're not done with a task until they hit a success criteria. And so as opposed to you having to sit there and prompt it, like keep trying, keep trying, keep trying on any given task. It knows to do that itself. This is not the first model to do this. Jipu AI's GLM 5.1, which came out two weeks ago now, I think, does that out of the box as well. And it can loop autonomously for about eight-ish hours. Mythos apparently can loop for significantly longer. What this means is it's not instant, but given a high-level criteria, like here's this banking website, find a way in. It will just keep trying until it achieves that thing. So these looping LLMs, what makes them powerful is that they can run very long horizon tasks. There's an organization called Meter, M-E that looks at how long can AI run a task successfully. And that number keeps going up. In December, Opus 4.5 was the winning model at like five and a half hours. And then Opus 4.6 came out in January. That's up to now 12 hours. And so the thinking is that Opus 5, which is what Mythos basically is, is probably around 24 hours. ([Time 0:00:48](https://share.snipd.com/snip/6b609574-6171-4268-8d16-9895211c107c)) - **Open Weights Make Advanced Models Widely Usable** - Open-weight models like GLM 5.1 can be downloaded and run locally, making potent capabilities widely accessible. - Christopher S. Penn warns that open models let motivated actors run untraceable long-running attacks if they have hardware. Transcript: Christopher S. Penn Yeah, they're going to get into most systems because there's going to be some little gap or vulnerability. But this is not the only model that can do that. And more importantly, GLM 5.1 is an open weights model, which means that you can download it today if you have the hardware and run it yourself untraceably and find your way into a bunch Of other things. ([Time 0:02:44](https://share.snipd.com/snip/60ab0be7-b63b-47fe-9f6e-5d3fc08e9a4a)) - **Patch Easy Holes And Add Defense Layers** - Reduce attack surfaces by patching common vulnerabilities and adding defense layers to shrink the Swiss cheese holes. - Christopher S. Penn recommends making the holes smaller and adding layers because AI navigates layered defenses effectively. Transcript: Christopher S. Penn Right. John Wall And well, then the thing that got me was there was talk about, okay, so they had an emergency meeting with the White House and Vice President Vance, Treasury Secretary, CEOs of Anthropic, OpenAI, Microsoft and Google. But the thing is, yes, you can say this is so dangerous in the wild, but I look at the flip side of, okay, now you just go patch all those bugs, you know? And so you could literally be in a place even a month from now where you can say, yeah, human hackers are no longer a thing. Like if you don't have an AI hacking tool behind you, you're not going to be able to get into anything. And so saying that the world is on fire and everybody's going to die is definitely irresponsible because there's, you know, so much more behind that and where it can go, but I'm not going To do my old man rant, but yeah. Christopher S. Penn Let's think about the people that are being briefed by the tech companies, the people who are being briefed in the white house and stuff like that are not what I would call the most technically Savvy people to begin with. I have a bunch of other words, but this is not a politics podcast. We'll save those words for at the bar. However, what the real challenge is for a lot of companies is not in like your state of the art systems. It is a social engineering and human engineering is a very real thing, right? There are uncounted numbers of AARP members whose password to their banking system is password 1234, right? And so these tools will be easily able to find that way. And the second and the much bigger problem, which is why there's national security implications, is that embedded systems, so these are things like the CPU in, say, a power substation. Some of them are like running Windows NT4 embedded from 20 years ago, and they've never been patched because the cost to go around every substation and rip out the hardware replace it With new hardware is astronomical. But all of that old hardware, things that control water valves, power relays, etc., are incredibly vulnerable. And with a tool like GLM 5.1 or Opus 4.6, even because there's a study showing Opus 4.6 can do this today you don't need mythos to do this even with smaller models you can say here's your Goal here's this you know this power station that's on the internet that's poorly secured go turn off all the lights like just brick the the hardware now you have mass chaos yeah that John Wall Is a huge, but like you said, water systems, power grids, like all of these, um, yeah, like you said, the, the hardware they're running is ancient and, you know, the last guy who knew how It, you know, operated, cashed in his COBOL machine, you know, five years ago and nothing's going on. So yeah, yeah. There's, there's all kinds of stuff, um, that, that are making a mess there. And then another thing is cybersecurity stocks taking a hit because of this. Like that, I totally see. Any tool from six months ago that's not AI powered, it's like you're going to war with a stick, you know? It's just like you're not going to get anything done. I would imagine that more of the cybersecurity firms are just going to have to make AI part of their suite. Is there any other thing that they can do about that? The big thing is patch the easy stuff. There's so many places. Christopher S. Penn Security is always like, they call it the Swiss cheese model, where if you take slices of Swiss cheese stacked up together, the holes are in different places. So theoretically, there's no clear line from one end of your block of cheese to the other, but there's only two slices or the holes are really big. You can get through pretty easily. So those multiple layers of defense, what's happening is that AI is very good at navigating those multiple layers of defense. So you either have to add more or make the holes smaller in the cheese. And that's really what they're talking about with a model like Mythos. Mythos is saying, hey, your cheese is actually mostly holes. ([Time 0:03:26](https://share.snipd.com/snip/eb39492e-dbe5-4aae-b51a-329cc6dd9629)) - **Legacy Embedded Systems Are National Security Risks** - The bigger national security risk is legacy embedded systems running decades-old, unpatched software controlling power and water. - Christopher S. Penn highlights devices running Windows NT4 embedded and unreachable hardware that can't be cheaply replaced. Transcript: Christopher S. Penn It is a social engineering and human engineering is a very real thing, right? There are uncounted numbers of AARP members whose password to their banking system is password 1234, right? And so these tools will be easily able to find that way. And the second and the much bigger problem, which is why there's national security implications, is that embedded systems, so these are things like the CPU in, say, a power substation. Some of them are like running Windows NT4 embedded from 20 years ago, and they've never been patched because the cost to go around every substation and rip out the hardware replace it With new hardware is astronomical. But all of that old hardware, things that control water valves, power relays, etc., are incredibly vulnerable. ([Time 0:04:35](https://share.snipd.com/snip/add8e3c3-71e2-4f2b-b95e-ef37a66a88f5)) - **Local Gemma 4 Heretic Delivered Accurate Dangerous Answers** - Christopher S. Penn tested Google's Gemma 4 locally using an uncensored Heretic variant and ran dangerous how-to queries for research. - He reports Gemma 4 answered correctly offline and used Tor/DuckDuckGo for web-sourced details when available. Transcript: Christopher S. Penn So Gemma 4 is Google's brand new open weights model. It is the sibling to Gemini. And the last version of Gemma was Gemma 3 that came out about a year and change ago now, maybe a little longer. Gemma 4 comes in a bunch of varieties and it is made for agentic AI. It is very good at tool handling. It is very good at all kinds of working in agent networks. It is very fast. There's like five varieties of it. There's Gemma 31B, 26B, A4B, 9B, 4B, E4B, E2B. And they scale from biggest to smallest because the smallest ones can run on an Android, right? So they have a local language model running on an Android or an iPhone and still be reasonably capable as long as it has access to tools like web search, for example. It is very fast. I've been testing it out on my machine. I'm using the MLX version for the Mac, and it is very, very fast. And I'm using a variant called Heretic, which is an uncensored version. I've been testing that out. One of the things, and I talked about this recently on LinkedIn, if you want truly guaranteed privacy, you have to use local models because anything in the cloud is somebody else's computer. So I feel perfectly safe asking all sorts of very dangerous questions to the local model that runs on my machine because it's not phoning home anywhere. And I asked it how to do some very bad things because I know the science behind them and the best way to do those very bad things. You would never want to type into anything on the internet. And Gemma 4's answers were correct out of the box for most of them. And when it used a secure web search tool through Tor and DuckDuckGo and stuff like that, it was like, yep, this is exactly how to do this very bad thing. This answers for educational use only was sort of its minor disclaimer. I'm like, that is pretty darn impressive. So it's available. It's free. If you have systems ([Time 0:10:32](https://share.snipd.com/snip/11ad22ad-1c14-4a20-ab99-10d7c8a11bd7)) - **Run Models Locally To Preserve Privacy** - Use local models for guaranteed privacy because cloud services run on other people's infrastructure. - Christopher S. Penn suggests running Gemma 4 locally on a Mac or via llama.cpp/LMStudio to avoid data phoning home. Transcript: Christopher S. Penn So I feel perfectly safe asking all sorts of very dangerous questions to the local model that runs on my machine because it's not phoning home anywhere. And I asked it how to do some very bad things because I know the science behind them and the best way to do those very bad things. You would never want to type into anything on the internet. And Gemma 4's answers were correct out of the box for most of them. And when it used a secure web search tool through Tor and DuckDuckGo and stuff like that, it was like, yep, this is exactly how to do this very bad thing. This answers for educational use only was sort of its minor disclaimer. I'm like, that is pretty darn impressive. So it's available. It's free. ([Time 0:11:51](https://share.snipd.com/snip/e0fce09b-11dd-47a1-8e27-17bdfdd9dc43)) - **Monetize By Packaging Prompts Into A Subscription** - Monetize content by extracting and repackaging prompts into a subscription prompt library delivered over time. - Christopher S. Penn describes converting 700 pieces into recurring two-issue weekly prompt deliveries plus prompt-to-skill plugins. Transcript: Christopher S. Penn So this is an ongoing topic of how the heck do I monetize my newsletter that has 300,000 subscribers? And what we came up with is on my newsletter, in my YouTube channel, in all over places, I'm constantly sharing stuff like here's how to do X, Y, or Z. And what we decided was let's take the corpus, all of it, like 700 pieces of content and put it through a language model and say, if there's a prompt in here, extract it. Because for good reason, a lot of people are not just going to sit and watch my YouTube channel every day. I wouldn't watch my own YouTube channel every day. And extract all the prompts, put them in the proper framework. And now it's a monthly subscription where you get two issues a week that have between one and five prompts in them from past episodes, past newsletters and things like that that you can Try out or better yet, one of the first things you get is a prompt to skill plugin for Claude or anti-gravity or Codex. You take the prompt and it'll turn it into a skill in your system with your tweaks that, you know, to make it truly yours. And then, you know, you find something like, oh, I want this to be the corporate strategy thing. It does a great job at converting that and making it available to you. So that's, that's what's going on with that. ([Time 0:13:12](https://share.snipd.com/snip/f732f69c-fa1e-4674-9c52-d61b459652bb)) - **Treat Marketing Survey Headlines With Skepticism** - Marketing studies often report consumer preferences without sample size or rigour, making headline claims unreliable. - John Wall points out a '90% want weekly texts' claim lacked visible sample size, undermining credibility. Transcript: John Wall It was funny. They said 90% of consumers want a brand text once a week or less. The only brand I know that does that is marketing over coffee. We guarantee you will get a text only once a month or so-ish, unless we're giving away amazing stuff. I don't really buy that. I mean, there's a lot of brands that text me every day even, and I ignore this stuff. And I was just thinking more about how we see these studies that are just, you know, it's a conglomeration of all this stuff going on. That doesn't necessarily mean that any of it is appropriate for your brand and what you are doing today. And again, it's kind of like, well, if 90% once a week or less, why haven't they just all unsubscribed? You know, like that makes no sense either. And of course the primary thing to look at was there was no number of the sample size. You know, we see so many of these marketing surveys where it's like, oh yeah, we got 65 people answered the survey. I'm just like, that's not even close to that. That's not how that works. The math is not mathing. ([Time 0:16:20](https://share.snipd.com/snip/28d3ed58-21d0-48f2-bd2d-b0bb001a6a94)) - **DIY YouTube Downloads Replace Premium For Travel** - Christopher S. Penn canceled YouTube Premium and built a personal downloader pipeline using Cloud Code and Jellyfin for ad-free offline viewing. - He schedules daily grabs from a marked playlist and stores videos locally to sync to devices for travel. Transcript: Christopher S. Penn I did end up canceling YouTube Premium because the price keeps going up. It's like $18.99, $19.99 a month now. And I'm like, okay, that's not that compelling for what you get. It's not wrapped into Google One. It's not wrapped into anything else. And so what I ended up doing was using Cloud Code to VibeCode, a downloader and YouTube manager for... So I have a playlist on YouTube, on my YouTube account now, where I just mark things for download. And then once a day, the script runs on my laptop, grabs all the videos, categorizes them, and puts them into... I'm using Jellyfin as my media server. Well, the other easy thing to do is you could just literally take those and push them to your iPad or whatever so that all the downloads are available on your phone and your iPad when you're Ready to hit the road. And so you can have all your queued stuff. So I am still not going to pay for YouTube premium. And now I have all the functionality of it, guaranteed no ads. ([Time 0:20:02](https://share.snipd.com/snip/4921842e-15bf-4769-b8b0-b1a69098f321))