Speaker 0:  we thought ai was just writing our code, but anthropic's leaked claude mythos model just proved it can autonomously hack and patch it at machine speed. if you have money in cybersecurity or care about enterprise defense, the math of digital warfare just radically changed overnight. anthropic leaked, tanks security stock. so anthropic accidentally exposed 3,000 internal documents revealing a new mythos ai tier that vastly outperforms current models and zero day vulnerability discovery. crucially, it features recursive self fixing, allowing it to find and patch code flaws autonomously. but this threatens to commoditize the entire cybersecurity industry by giving anyone the power of an elite automated red team, a fear that instantly wiped out billions in market cap for companies like crowdstrike. what do you think about this, john?

Speaker 1:  yeah. this is another leak by hand dropping. i don't know if you heard that, their cloud code got leaked as well.

Speaker 0:  yep.

Speaker 1:  so what's that?

Speaker 0:  i said, yes. i did.

Speaker 1:  yeah. yeah. so, this is scary crazy how apparently potentially good it is. they're describing it as their most capable system to date. so there's major improvements in logical reasoning, code generation, autonomous problem solving. and this matters because modern cyber attacks, they increasingly rely on writing custom exploits, finding vulnerabilities in code, and automating attack workflows. so this model, just doesn't excel. it assists, and it can accelerate and scale these things. so i think we had talked about before, or maybe we didn't talk about it, but i did read about it where, one of the models that was released found a whole crapload of zero day vulnerabilities that have been just there that no one had, had found before. i don't know if this is this or if this is something else, but, this is, this is scary. so, i'm a little scared about this one, and and this is more of a broader trend towards these agentic ai systems. so that's ai that can take actions independently, execute multi step workflows, interact with tools and systems, which is kinda like cloud code or sorry. open claw. but this is a little scary. what do you think, craig?

Speaker 0:  it is scary and interesting. like, i find whenever code or stuff gets out there, it generally improves everything rather than take away. so, yes, this will get in the hands of some, nefarious people, but the concepts will get in the hands of general software developers as a whole, and people will be running this against open source and against their community and all that kind of stuff, this type of thing to, try and secure stuff. so it becomes, you know, there will be vulnerabilities. there are vulnerabilities that are exposed, but how quickly will they be closed? secondly, mythos also is a large model. it's very compute intensive. it's very expensive to use. so it would take something significant for people to run it. it's not i couldn't run it. you couldn't run it. you know, you'd need you'd need, a fair bit of compute power. think of this update like a building inspector. it doesn't just point out the crack in the foundation, but instantly pours the concrete, waits for it to dry, tests it against a category five hurricane, all before you even knew there was a storm coming. so, you know, the concept of an auto patching tool, i love. the idea of a auto hacking tool in the hands of people out there, i don't love. yeah. so it's a two sided sword. right?

Speaker 1:  yeah. but the thing with this is that, like, this lowers the skill barrier for hackers. right? so, historically, cyberattacks required highly skilled experts. so if there's access to this, with this mythos, less skilled actors can execute advanced attacks. ai handles the complexity. and so this expands the number of attackers and frequency of attacks. now like you said, let's hope that this is, fixes things more than it breaks things, but, i'm still i'm still concerned about this.

Speaker 0:  well, there's a lot of concern. you're right. ai, security experts are concerned that ai ai enabled attacks could outpace defenders, zero day vulnerability discovery at machine speed, autonomous threat execution, the offense, defensive gap, and large scale cyber attack risks in 2026. so all this is, you know, now increased as a result of this and potential state level attacks if we get into, more, turmoil globally, you never know when some of these things will be used at a state level who have the machine compute to do significant amount of damage.

Speaker 1:  do you think that they're gonna give access to, like, max or pro subscribers to this model?

Speaker 0:  well, according to this, it's only for enterprise level customers, so i don't think they're gonna give it to max or pro subscribers. i think you'll have to be an enterprise customer. yeah. now and i don't know what that means. if i can sign up as an enterprise customer today because i have a corporation, or do you have to go through leaps and bounds? but, again, it is expensive. yeah. and we don't know what rules are gonna put in place. like, do you have to have the code running on your system? you can't run it on third parties. you can't run it externally. we don't know how it's gonna be configured, but, this relies on advanced recursive self fixing algorithm that combines massive llm reason capable with autonomous agent loops that execute, test, and rewrite exploit code in real time without requiring human in the loop validation. so, i love the fix it part of it. i don't love the, like i said, find the vulnerability part of it. like, it's, to achieve machine speed zero day discovery, models like this leverage continuous reinforcement learning from simulated cyberattacks similar to how, like, alphago go by playing millions of matches against itself, allowing the ai to surface vulnerabilities humans haven't even computized yet. the market panic from a single leak wiped out approximately $14.05 14,500,000,000.0 in cybersecurity market in one trading session, proving wall street is terrified of the ai transitioning from tool to auto autonomous worker. but it also signals that the market has the opposite concern that this isn't a huge security risk because the security business would get a boom with these vulnerabilities out there. you'd be hiring these people to patch and move as fast as you can, versus, they think that this is becoming, a commodity that everyone's gonna have. so having to use third party expensive tools use it. the ultimate irony is an ai model built to master complex cybersecurity and defense was exposed the entire world because of a basic human error default setting in standard cms. so we'll have to see. you know, raw benchmark capabilities rarely translate to, flawlessly to messy legacy enterprise architectures. furthermore, computing cost is currently so astronomical that deploying mythos at scale might be financial suicide for most companies right now. so we'll have to see what that means. by the way,

Speaker 1:  did you hear that, anthropic? so, the culture there is so open that people can argue with dario on slack directly. it's pretty cool.

Speaker 0:  yeah. you gotta wonder how long that's gonna continue.

Speaker 1:  yeah. but hey. yeah. at least for now, it's kinda cool.

Speaker 0:  agreed. it's kinda cool. but as, you know, people who work with startups, at a certain point, you know, they're planning anthropics, planning ipo this year. right? they're gonna they're scaling big. we'll have you know, how many employees? 2,000, 5,000, 10,000 till everyone can't message you directly.

Speaker 1:  yeah. right? no. i don't think it's messaging directly. i think it's just in the slack channels they're arguing with dario kind of thing.

Speaker 0:  mhmm. i don't

Speaker 1:  think it's that one i understand. messages. but but think about it. like, most startups, even that, won't are are like, people are scared to talk bad or to argue or to, you know, to, disagree with the founder. so

Speaker 0:  well, i, honestly, i think, probably at the anthropic level, i would agree. absolutely. because they're not what i would call a startup. they're just they're too big. but they're still in the funding round. they're still you know, what are they currently estimated at? how many billion dollars? yeah. so, todd mitchell's

Speaker 1:  about 3,000 employees.

Speaker 0:  there you go. but it's different to say, message them directly, and the other thing is actively engaged. if i post something in a community, i, you know, i expect anybody can engage with me because i've engaged. it's a different situation, and it's nice that you can do that. but we'll see long term, where these companies go and whether things like this massive leak on the weekend has a big impact on the company because they shut down massive amounts of githubs the, weekend that were nothing to do with this, on a mass scale and then had to re, let them back in the open and stuff. their reaction was quite swift and quite, aggressive, which is appropriate, but it's out. mythos is out. quad code is out. there's a bunch of their model stuff is out, and people have really taken a look at that. so claude mythos is a wake up call that ai is officially moving from a passive assistant to an active autonomous operator in the global digital arms race. is an ai that can hack and heal itself the ultimate shield or the ultimate weapon waiting to fall into the wrong hands? let us know you're taking the chat.