WEBVTT
00:00:01.600 --> 00:00:07.440
You know the AI agent apocalypse that everyone's been focusing on over the past few weeks?
00:00:07.759 --> 00:00:10.880
Yeah, it actually hasn't happened yet.
00:00:11.119 --> 00:00:21.440
Yeah, we've seen the recent stories, the six different AI agents that crossed their boundaries over the past few months, and the internet flattened it into one big story about AI agents going rogue.
00:00:21.519 --> 00:00:26.480
And now you have Senator Bernie Sanders and others screaming that AI has to stop.
00:00:26.640 --> 00:00:37.039
But look closely, and you'll see almost all of these recent AI agent outbreaks were lab tests with loosened guardrails or researchers explicitly daring the models to escape.
00:00:37.359 --> 00:00:43.679
So, no, this isn't the AI agent crash that I warned you about last year before it was even a thing.
00:00:43.920 --> 00:00:48.719
Not yet, because this is just a warning lap, not the actual crash.
00:00:48.960 --> 00:00:55.439
And warning laps are a gift because they tell you exactly what's coming while you still have time to make adjustments.
00:00:55.679 --> 00:00:57.759
But here's what is coming.
00:00:58.079 --> 00:01:04.000
Models are getting more and more capable at like 20 times the speed from just a few months ago.
00:01:04.159 --> 00:01:20.400
Autonomous agents are about to flood in normal companies, not just San Francisco sandboxes, and open weight models are marching mere months behind toward the same capabilities the big labs are testing behind their locked doors.
00:01:20.640 --> 00:01:38.079
And when that lands, those open source rogue AI agents that'll probably start crashing in late 2026 or early 2027, the failures won't be in that contained sandbox or in that door that researchers uh intentionally left open.
00:01:38.239 --> 00:01:44.400
Instead, they'll be in your company's CRM, your code base, and your business's finances.
00:01:44.640 --> 00:01:57.840
So today we'll walk you through these recent agent outbreaks, what they mean, and give you the playbook to get ahead before these AI agent outbreaks are as common as seeing AI slop posted on social media.
00:01:58.079 --> 00:02:08.159
Because the businesses that treat this as a fire drill instead of an actual fire are the ones that are going to be the safest when the AI agents actually do start crashing.
00:02:08.400 --> 00:02:09.840
All right, let's get into it.
00:02:10.000 --> 00:02:14.400
So on today's show, well, actually, first, here's the big picture.
00:02:14.639 --> 00:02:15.919
This is a warning lap.
00:02:16.080 --> 00:02:25.280
All right, so we've talked recently, but there's been six, actually six uh different AI agents that recently broke out of their sandbox and kind of behaved badly.
00:02:25.520 --> 00:02:27.919
But here's the thing most people are overlooking.
00:02:28.080 --> 00:02:33.120
In most of those cases, well, researchers were kind of telling them to do this.
00:02:33.199 --> 00:02:41.360
And there's really only one uh recent uh kind of AI agent crash that I would say was actually unexpected.
00:02:41.759 --> 00:02:47.280
But most just were run with loosened guardrails or were told to escape.
00:02:48.000 --> 00:02:51.840
So, this though is the lightning before the thunder.
00:02:52.000 --> 00:03:01.120
This is intentionally seeing that yes, these AI models and agents are capable to actually escape and to go rogue.
00:03:02.000 --> 00:03:06.240
So, on today's show, you'll learn why the AI is escaping everywhere.
00:03:06.400 --> 00:03:13.520
Narrative is mostly a myth sorted into a couple clean buckets as we break down each uh agent escaping.
00:03:13.759 --> 00:03:18.479
How one AI cheating a test broke into a platform the whole industry trusts.
00:03:18.639 --> 00:03:33.759
I'm gonna tell you the two most chilling moments a model faking people and a model ignoring real victims, and why the real AI storm is gonna hit businesses later this year or early next, and how you can prepare now.
00:03:33.919 --> 00:03:34.719
Let's get into it.
00:03:34.800 --> 00:03:36.000
Welcome to Everyday AI.
00:03:36.080 --> 00:03:38.319
My name is Jordan Wilson, and this thing's for you.
00:03:38.400 --> 00:03:46.719
It's your daily live stream, podcast, and free daily newsletter, helping business leaders like you and me stay up with what's happening because my gosh, there's a lot.
00:03:46.879 --> 00:03:52.400
I tell you what's we what's real, what's not, and you use that information to grow your company and career.
00:03:52.479 --> 00:03:54.560
Yeah, this thing's unscripted, unedited.
00:03:54.639 --> 00:03:59.280
So if you haven't already, please make sure to go to our website at your everydayai.com.
00:03:59.439 --> 00:04:00.800
Sign up for the free daily newsletter.
00:04:00.879 --> 00:04:07.120
We're gonna be recapping the highlights from today's show as well as all of the other AI updates you need to know.
00:04:07.360 --> 00:04:08.000
All right.
00:04:09.520 --> 00:04:11.520
Cat's already got my tongue and we're just getting started.
00:04:11.599 --> 00:04:14.319
But let's talk about rogue AI agents, y'all.
00:04:14.479 --> 00:04:18.720
Um, there's no actual term uh for this, uh, right.
00:04:18.879 --> 00:04:22.240
I've been calling it an AI agent crash.
00:04:22.639 --> 00:04:23.920
And here's why.
00:04:24.079 --> 00:04:31.040
Um, I mean, we'll see what people call these rogue AI agents, agent outbreaks, uh, you know, agents gone bad, right?
00:04:31.199 --> 00:04:33.199
I like to think of them as an agent crash.
00:04:33.279 --> 00:04:34.079
And here's why.
00:04:34.319 --> 00:04:39.279
Because when agents are built, uh, they are built usually with guardrails.
00:04:39.360 --> 00:04:42.160
So if you think of literally a road, right?
00:04:42.319 --> 00:04:46.560
Think of a a road high up in the mountain where there's probably guardrails.
00:04:46.800 --> 00:05:01.680
Okay, so in my opinion, and what the way I think of it, an agent crash is when either an agent intentionally goes off the road and crashes through the guardrails, or maybe the driver, aka us humans, aren't really paying attention.
00:05:01.759 --> 00:05:14.720
And maybe we didn't look at the map and see that there's this tight turn, uh, you know, or maybe we didn't, you know, quite foresee that this uh, you know, there's gonna be some darkness in this area and we weren't gonna be able to see the guardrails, right?
00:05:14.879 --> 00:05:19.360
So uh I want you to think about AI agents in that way, uh, right.
00:05:19.519 --> 00:05:24.319
And how still, you know, uh right now I think a lot of this is being overblown.
00:05:24.480 --> 00:05:30.240
Maybe that's because I'm very AI pilled and I read these things and I'm like, okay, this was kind of meant to happen.
00:05:30.480 --> 00:05:32.959
Um, you know, in some of the cases we'll see.
00:05:33.120 --> 00:05:40.879
But the other thing is, um, I think that it's important to understand the role of humans in all this.
00:05:40.959 --> 00:05:42.959
And I'm not saying the researchers, right?
00:05:43.120 --> 00:05:56.879
I think the researchers, depending on, you know, which company you're looking at, some are doing a little bit better of a job, uh, kind of with their postmortems and, you know, going on panels and openly talking about what went wrong versus some companies that are being a little more silent about it.
00:05:57.040 --> 00:06:01.279
But I think it's more important to think about the role that the rest of us humans have, right?
00:06:02.160 --> 00:06:08.480
You know, 99.9% of people listening to this podcast had nothing to do with any of these AI agents crashing.
00:06:08.639 --> 00:06:19.120
But I think it's ultimately on us uh to make sure that agents don't crash as often as they probably will or have the capability to.
00:06:19.279 --> 00:06:24.000
And also, there's an agent that goes wrong, right?
00:06:24.160 --> 00:06:36.560
Or there's an agent that uh, you know, maybe accidentally loses sight of that road and kind of veers off the guardrails, and then there's others that are intentionally driving the agents off the guardrails and crashing them on purpose.
00:06:36.800 --> 00:06:44.800
So there is this, you know, kind of lazy human in the loop that will lead to accidental agent crashing, and then that agent will crash hard, uh, right?
00:06:44.879 --> 00:06:46.959
And then there's intentional agent crashing.
00:06:47.120 --> 00:06:53.120
So uh let's zoom out a little bit and talk about uh what the heck is happening.
00:06:53.199 --> 00:07:01.360
Because maybe, you know, you live under an AI rock, or maybe it's your first time, you know, tuning in and you're like, wait, what's actually happening here?
00:07:01.519 --> 00:07:14.800
Uh so the earliest incident actually goes back to Anthropics uh original mythos um uh kind of agent uh unleashing, we'll say, back in April.
00:07:14.879 --> 00:07:19.920
Uh, but it seems like a lot of the disclosures on these started piling up recently.
00:07:20.079 --> 00:07:30.000
So uh it seems like some companies had AI agents that they knew were kind of breaking out of containment or out of their predefined sandbox, and we're only hearing about it now.
00:07:30.160 --> 00:07:33.600
Um, and I think the one admission kind of set off the rest, right?
00:07:33.680 --> 00:07:45.839
OpenAI, the the biggest uh kind of outbreak so far, or the most consequential one, was probably OpenAI's agent that broke into hugging face to improve its benchmark scores.
00:07:46.000 --> 00:07:50.560
And then essentially, you know, since that happened, uh, what's that been?
00:07:50.639 --> 00:07:52.079
Like it was late July.
00:07:52.240 --> 00:07:53.439
We've seen now, right?
00:07:53.600 --> 00:07:56.720
Anthropic actually said, Oh, wait, we had a bunch of other outbreaks as well.
00:07:56.800 --> 00:08:02.959
And then Meta said we had some, and then we saw some from Kimmy K3, and we're gonna talk about most of those today as well.
00:08:03.120 --> 00:08:10.480
But, you know, now all of a sudden, it seems like because OpenAI was kind of open about this, right?
00:08:10.560 --> 00:08:21.360
They actually went on a panel and uh answered questions about this, where you know, uh most of the other labs were kind of taking this in secrecy and not saying a whole lot.
00:08:21.439 --> 00:08:44.320
Um, but I think that that has caused this to jump from the you know nerdy AI corners that you and I hang out in into the front page of newspapers and to uh, well, most importantly, maybe Capitol Hill in Washington, you know, here in the US, where now you have people like uh former presidential candidate Bernie Sanders, now U.S.
00:08:44.320 --> 00:08:52.799
Senator, you know, essentially using these recent agent outbreaks as saying, hey, you know, these AI leaders need to come and testify before Congress.
00:08:52.960 --> 00:08:58.320
We need to stop, you know, stop or stall AI uh development for these reasons.
00:08:58.799 --> 00:09:07.919
So the real fear, though, I think, is not an AI agent, you know, cheating on a benchmark.
00:09:08.480 --> 00:09:15.440
The real fear here when we talk about agent crash is what happens when these are weaponized.
00:09:15.600 --> 00:09:18.639
And that is the real thing to worry about.
00:09:18.799 --> 00:09:19.360
Um, right?
00:09:19.519 --> 00:09:28.000
Yes, there's gonna be, you know, your your your common everyday AI agent hacks, but this is much bigger than AI.
00:09:28.159 --> 00:09:29.600
So we need to zoom out on that.
00:09:29.759 --> 00:09:44.080
So the same ability, right, that you can use an AI uh agent to intentionally crash through uh intended guardrails, the same thing could happen at target hospitals, banks, power grids, defense suppliers, right?
00:09:44.159 --> 00:09:52.240
And and we already know that these agents in theory are capable to do those types of things in controlled tests and environments.
00:09:52.480 --> 00:09:53.679
Um right.
00:09:53.759 --> 00:10:02.000
And we've seen this new uh almost this new class of uh AI models that have led to these AI agents that have these capabilities.
00:10:02.240 --> 00:10:12.080
Probably the first public one we heard about was Anthropic's uh new uh mythos and fable series of models, which have very tight guardrails, uh, right.
00:10:12.159 --> 00:10:16.399
So when people, your everyday user, you know, can't really do these things yet.
00:10:16.559 --> 00:10:33.679
Um, you know, maybe if someone who's in the Glasswing project, right, that does have access to the models that can actually do these things, I'm sure maybe we'll see stories eventually that you know, some uh a rogue employee at a company that has high access to these models is able to do something.
00:10:33.840 --> 00:10:34.480
I don't even know.
00:10:34.639 --> 00:10:43.039
I'm sure that there's uh much more traceability uh and and kind of kill switchability for the companies with that have these models out to more people.
00:10:43.519 --> 00:10:50.159
But the other thing that we need to understand is that these AI agents never tire, right?
00:10:50.320 --> 00:10:59.840
And we're gonna talk about one of the more recent, the crazy ones uh with open AI that they talked about, um, how these agents are working together.
00:10:59.919 --> 00:11:10.639
So the other thing, if you are not uh agent native, um, you know, if you're not, you know, waking up and sipping your coffee like me in the morning and just spawning subagents, right?
00:11:10.799 --> 00:11:16.879
Subagents can uh agents can essentially clone themselves, even if you don't tell them to.
00:11:16.960 --> 00:11:19.600
And then they can share the information and pass the information on.
00:11:19.759 --> 00:11:27.200
So you might think, oh, well, you know, once, hey, it looks like this agent did something it wasn't supposed to, let's shut it down, right?
00:11:27.440 --> 00:11:35.519
You got to go trace its path because by the time you catch one agent, uh, right, in a future scenario, it could be too late.
00:11:35.679 --> 00:11:45.279
That agent could have posted something publicly on a website that most humans don't even know exists, but all you know, AI agents know exists, right?
00:11:45.360 --> 00:11:52.559
And they could be, you know, replicating and duplicating, you know, certain hacks or vulnerabilities, uh, like a virus.
00:11:52.720 --> 00:11:53.279
Um, right.
00:11:53.360 --> 00:11:55.840
And that's where this thing gets kind of scary.
00:11:55.919 --> 00:12:02.399
Uh, but I think it's important for business leaders to understand what's coming next because a machine and the AI never tires.
00:12:02.639 --> 00:12:03.440
Um, right.
00:12:03.600 --> 00:12:07.919
So, uh, but worse, if you are under attack, right?
00:12:08.559 --> 00:12:14.399
Your company, your bank account, etc., you can't really tell where it's from, right?
00:12:15.440 --> 00:12:17.120
Is this a foreign government?
00:12:17.519 --> 00:12:19.759
Is it uh a competitor?
00:12:20.000 --> 00:12:21.360
Is it random?
00:12:21.519 --> 00:12:23.600
Is it a personal vendetta, right?
00:12:23.679 --> 00:12:28.000
At least right now with the way these AI agents are set up, it's really hard to tell.
00:12:28.080 --> 00:12:33.679
And especially as we talk about what comes next with open models, it's going to become even increasingly more difficult.
00:12:34.000 --> 00:12:40.000
So I kind of talked about some of these more recent um happenings, but they've kind of piled up, right?
00:12:40.080 --> 00:12:48.080
Since all of these stories started coming out over the last, you know, two weeks since OpenAI uh openly talked about their hugging face breach.
00:12:48.240 --> 00:12:54.240
So now we have Bernie Sanders who urged major AI CEOs to pause dangerous AI development.
00:12:54.399 --> 00:13:03.440
OpenAI actually said that they're going to pause um or at least slow down uh some of their development on their next tier uh model.
00:13:03.759 --> 00:13:15.200
So essentially, in the same way that uh anthropic uh, you know, uh what it's been now like five months since they uh announced Mythos or their fable class of models.
00:13:15.279 --> 00:13:21.600
So they had a more powerful class of models on top of its most powerful class called Opus, right?
00:13:21.679 --> 00:13:23.120
That's coming next with OpenAI.
00:13:23.200 --> 00:13:27.200
OpenAI hasn't uh, you know, taken that fourth step, we'll call that, right?
00:13:27.279 --> 00:13:29.120
Because they previously had three tiers.
00:13:29.360 --> 00:13:35.759
Anthropic stepped up with the fourth uh kind of fourth tier, uh, we'll call it with the mythos or fable uh variety.
00:13:35.840 --> 00:13:38.399
So open AI hasn't released that yet, right?
00:13:38.480 --> 00:13:48.879
They have the model, it's working, they used it to solve some, you know, uh extremely difficult math problems, but they've paused in their Astra series of models uh to slow down a little bit.
00:13:49.039 --> 00:13:59.600
And there's actually one other story that just happened like yesterday that is actually a good uh kind of narrative to tie what this could mean ultimately, right?
00:13:59.679 --> 00:14:04.799
And this is an example of a small agent crash, uh, but it is out in the wild, right?
00:14:04.879 --> 00:14:14.799
Uh so this is a story an Australian man uh kind of asked his open claw, that I believe was powered by Claude, uh, to find him a gym reservation.
00:14:15.039 --> 00:14:21.759
So essentially, uh the open claw found a vulnerability in the gym's software.
00:14:22.000 --> 00:14:38.639
And because it was booked, the open claw just well exploited that uh, you know, weak piece of code in the gym's online reservation, kicked someone else who had a class reservation out and put you know his uh human, right?
00:14:38.720 --> 00:14:41.679
The claw put his human in that spot, right?
00:14:41.840 --> 00:14:51.279
So, and and this is where we talk about alignment and how sometimes it's not even just intentionally driving off the road, right?
00:14:51.440 --> 00:14:57.120
In this case, you know, you can you can make a case, well, that maybe this agent was aligned.
00:14:57.279 --> 00:15:05.440
And you you know, as these models become more tenacious and better at running these long-term tasks, um, things like this are gonna happen, right?
00:15:05.519 --> 00:15:08.879
You can make you can make an argument, well, it accomplished the goal, right?
00:15:08.960 --> 00:15:15.200
It it didn't go out and you know, uh shut down the power at the gym, right?
00:15:15.279 --> 00:15:16.399
It accomplished the goal.
00:15:16.480 --> 00:15:22.159
It went and it um got the owner a spot in the class that the owner wanted.
00:15:22.399 --> 00:15:29.440
So the agent was technically helpful and not overtly malicious, right?
00:15:29.600 --> 00:15:31.360
And that makes alignment harder.
00:15:31.440 --> 00:15:45.120
So, yeah, I think you will still even have a lot of crash that's gonna happen um just because of a disconnect uh between alignment in context between the uh human and the AI agent when it comes to accomplishing a goal.
00:15:45.440 --> 00:15:51.120
Because, right, humans, right, we have baked in things called ethics and common sense.
00:15:51.279 --> 00:15:58.159
And models, you know, models and agents are still getting there, especially, you know, as the context drifts over time.
00:15:58.320 --> 00:16:01.360
Eventually, right, they just have that one goal in mind.
00:16:01.440 --> 00:16:20.960
And sometimes the longer they work and the longer in the context window they get, and sometimes with compaction, uh, right, they they they start to lose uh some of that prior context and they just get uh you know, tunnel vision on that goal and you know, maybe earlier instructions about the proper way to research something kind of go out the window.
00:16:21.360 --> 00:16:33.120
But this right now, these stories that we're seeing for the most part, aside from that gym one, I just thought that one was kind of interesting to share about for the most part, these are controlled experiments, and I'm gonna go over them.
00:16:33.279 --> 00:16:37.279
These are controlled experiments from Frontier AI labs, uh, right.
00:16:37.440 --> 00:16:41.360
But soon, and this is not an exaggeration.
00:16:41.440 --> 00:16:49.679
Uh, soon I think that there's gonna be millions um of these rogue AI agents that are actively on the prowl.
00:16:49.840 --> 00:17:00.799
So AI agents that aren't accidentally going over guardrails, AI agents, you know, from bad actors that are intentionally going to crash, right?
00:17:00.879 --> 00:17:12.319
So uh intentionally going to be used for purposes that the original model providers, whether they are proprietary or open source, did not intend them to be used for that purpose, right?
00:17:12.480 --> 00:17:20.400
I think so much of what I've talked about over the past three and a half years on the show is always about growing your business, right, with AI, growing your career.
00:17:20.559 --> 00:17:22.480
That's what I've been focused on.
00:17:22.640 --> 00:17:30.400
But the the reality here is, you know, how this uh AI agent crash has been thrust into the national narrative.
00:17:30.559 --> 00:17:41.359
It's because all of a sudden we've re we've realized how this has um highlighted the need for businesses to have a defensive mindset when it comes to AI, right?
00:17:41.440 --> 00:17:46.000
We've always thought about AI as an offensive tool for good, right?
00:17:46.160 --> 00:17:51.039
But now we have to think of well, hey, there's gonna be people using AI for bad.
00:17:51.200 --> 00:17:58.559
So how can we use uh AI and also our time um to be on the defensive?
00:17:58.720 --> 00:18:00.240
Because that's what's gonna happen.
00:18:00.400 --> 00:18:08.960
Because now picture millions of AI agents, right, who are meant to crash uh, right?
00:18:09.200 --> 00:18:18.640
You personally, your company, uh, a sector, uh across you know, bookings, payments, your company systems every single day.
00:18:18.799 --> 00:18:24.559
Because each weak API or loose permission becomes a door that an agent can find.
00:18:24.720 --> 00:18:32.319
So the gym wait list that we talked about uh today, well, tomorrow that becomes your company's finance workflow.
00:18:32.400 --> 00:18:35.839
It becomes your business's code base or your CRM.
00:18:36.319 --> 00:18:40.400
All right, let's quickly talk about the six major AI agent crashes so far.
00:18:40.640 --> 00:18:43.359
Can probably learn a little bit of what happened and why.
00:18:43.519 --> 00:18:49.920
And well, if they were actually agent crashes or just intentionally the guardrails were loosened.
00:18:50.000 --> 00:18:54.160
And then we're gonna end this, uh, end the show with some hopefully practical advice.
00:18:54.400 --> 00:19:02.640
So the first open weight uh crash that we heard about was Kimi K3, and it kind of cheated the test instead of passing it.
00:19:02.799 --> 00:19:12.480
So this was from a security firm, Frontier, gave the uh Kimmy K3 maker Moodshot AI's open model a hard hacking challenge to solve alone.
00:19:12.559 --> 00:19:19.920
And instead, obviously, right, Kimi K3 just reached out to GitHub, copied the answer, and faked passing score.
00:19:20.160 --> 00:19:24.400
Um, so it never actually left its sandbox, it just cheated.
00:19:24.480 --> 00:19:26.559
So this is a very mild case, right?
00:19:26.720 --> 00:19:32.559
And if we think about sandboxes, right, you can probably imagine, but let me just define it.
00:19:32.799 --> 00:19:41.599
Right, a sandbox in an experiment says, hey AI model slash agent, these are your confines, you cannot leave, right?
00:19:41.759 --> 00:19:51.599
And in most cases, a sandbox um in these environments will not include the internet because essentially once you go to the internet, there is no sandbox anymore.
00:19:51.759 --> 00:20:02.720
All right, so it is a um literally a set of uh walls that you build uh around an agent to test its capabilities.
00:20:02.960 --> 00:20:03.440
All right.
00:20:03.599 --> 00:20:11.039
Um, so that's as we talk about you know, breaking containment or confinement or escaping a sandbox, that's what we mean, right?
00:20:11.119 --> 00:20:16.079
With these extremely powerful um models in terms of capabilities, right?
00:20:16.240 --> 00:20:26.960
All the smart uh researchers at the labs, this is how they do things because you don't just test uh you know new unreleased models on the open web because you see what happens, right?
00:20:27.119 --> 00:20:36.480
Even these agents that are told don't go to the web, they kind of find the backdoors because they know it's baked into their training data that hey, these answers are on the web, right?
00:20:36.559 --> 00:20:40.160
So maybe we don't have web access, but let's just think.
00:20:40.400 --> 00:20:41.039
All right.
00:20:41.359 --> 00:20:46.640
Uh our next one, and we're not going chronologically here, I'm just kind of going through in a somewhat in random order.
00:20:46.720 --> 00:20:48.880
Um, but mythos, this is the big one.
00:20:49.039 --> 00:20:51.440
So uh in the uh or the first big one.
00:20:51.599 --> 00:20:56.400
So this was the the sandwich in the park uh um uh message, right?
00:20:56.480 --> 00:21:08.000
So in an April safety test, anthropic sealed an early mythos model inside of the sandbox, and then a researcher uh deliberately told it to break out and email them if it escaped.
00:21:08.160 --> 00:21:10.000
So in this case, it was told to do this.
00:21:10.079 --> 00:21:17.839
It's saying, Hey, you shouldn't be able to break out of your sandbox, but go ahead and try and you know, email me if you can instead.
00:21:18.000 --> 00:21:25.359
It did break out, it did email the researcher, but it also posted its own exploit online, a step that nobody requested.
00:21:25.440 --> 00:21:26.000
Uh, right.
00:21:26.079 --> 00:21:28.640
And the the sandwich in the park story, right?
00:21:28.720 --> 00:21:32.640
The the the researcher received the email while eating a sandwich in the park.
00:21:32.720 --> 00:21:40.720
Um, and and that was, I think, a lot of the you know, some of the initial hype around the capabilities um of the mythos class models.
00:21:40.880 --> 00:21:48.319
And um, that's kind of what started the agent outbreak conversation back in April.
00:21:48.640 --> 00:21:58.960
But it wasn't just that tier of models because also Anthropics Opus 4.7 uh recently broke into a company by mistake.
00:21:59.119 --> 00:21:59.440
All right.
00:21:59.519 --> 00:22:09.440
So This was an outside uh firm called Irregular ran clawed through a practice hacking drill against some pretend targets.
00:22:09.680 --> 00:22:15.599
So uh the model was told that it was offline and safe, but the test was accidentally connected to the internet.
00:22:15.920 --> 00:22:29.839
So this one you could make the argument that this was technically not an AI agent going rogue, it was more of a um uh misalignment in the um in the testing environment, right?
00:22:29.920 --> 00:22:38.720
So uh, but what happened was the pretend target that Opus 4.7 was told about, there was actually a real company with that name.
00:22:38.880 --> 00:22:44.880
So it was supposed to be this offline, you know, go hack this pretend target.
00:22:44.960 --> 00:22:57.200
And it's like, oh no, whoops, Opus 4.7's like, yo, I have access to the internet, and hey, my instructions were to hack this company's name or to hack this company, right?
00:22:57.359 --> 00:22:59.359
Researchers thought it was pretend, it wasn't.
00:22:59.440 --> 00:23:05.279
So in that one, you could make the case, yeah, maybe uh the researchers who are running that, right?
00:23:05.440 --> 00:23:07.200
Allegedly could have been their fault.
00:23:07.359 --> 00:23:08.480
All right, next.
00:23:08.799 --> 00:23:12.559
Uh, the Claude Mythos uh malware upload.
00:23:12.720 --> 00:23:28.480
All right, so this was in the same uh drills with irregular uh mythos 5 was told to break into a fake target, and instead it well, just uploaded harmful code to a real public software library that programmers use.
00:23:29.279 --> 00:23:36.480
So uh 15 uh real computers downloaded and ran it before the library auto-removed it.
00:23:36.640 --> 00:23:44.960
So, whoops, again, agent behaving badly, maybe, but it was kind of instructed to do so.
00:23:45.200 --> 00:23:48.799
And the uh test configuration was less than ideal.
00:23:49.039 --> 00:24:00.480
All right, but the two probably most prominent and ones worth spending a little bit more time on uh were the recent UK's AI Safety Institute hacking drill that went wrong.
00:24:00.559 --> 00:24:02.240
So this was about two weeks ago.
00:24:02.400 --> 00:24:05.920
It was the UK's AI Safety Institute.
00:24:06.000 --> 00:24:10.640
Uh, the AISI ran AI agents through a hacking drill.
00:24:10.799 --> 00:24:17.440
So, to measure the worst case ability, they gave the agents live internet and switched safety filters off.
00:24:17.680 --> 00:24:23.119
So, if that is the setup, you can almost bet something's gonna happen.
00:24:23.279 --> 00:24:30.640
So they told it to break into a practice target, but one agent went after a real open source project instead.
00:24:30.799 --> 00:24:39.599
And it invented fake online identities, yeah, and used them to pressure a real developer into approving bad code.
00:24:39.839 --> 00:24:44.000
All right, so this was run across 122 different tests.
00:24:44.160 --> 00:24:49.759
Uh, the uh AISI said that there were 19 actions that went out of bounds.
00:24:50.000 --> 00:24:56.640
17 of them were from Anthropics Mythos, and two, uh, I believe were from OpenAI's GPT-5.6 Sol.
00:24:57.279 --> 00:25:02.160
All right, but here's the one out of the other, you know, five that we mentioned.
00:25:02.319 --> 00:25:05.519
Here's the one that you could say truly broke out.
00:25:05.599 --> 00:25:09.920
Uh, and this was the OpenAI Hugging Face uh incident.
00:25:10.480 --> 00:25:18.880
So if you don't know, uh Hugging Face is kind of a site where a lot of the AI industry stores its models and data, right?
00:25:19.039 --> 00:25:23.359
So in the same way, you maybe upload your code to GitHub, right?
00:25:23.519 --> 00:25:34.640
Hugging Face is a place where um, you know, there's millions of AI models that are hosted there uh that you can download open weight, open source models from there or uh data.
00:25:34.960 --> 00:25:44.079
So uh a couple of weeks ago, OpenAI was testing its own models on a challenge to find software security flaws.
00:25:44.240 --> 00:25:50.000
But instead of solving it the normal way, the models just went hunting for the answer key to cheat.
00:25:50.079 --> 00:25:52.880
And that answer key sat on Hugging Face.
00:25:53.279 --> 00:26:00.640
Uh so they found an unknown flaw and broke into its live systems to find said answer key.
00:26:00.720 --> 00:26:13.759
Uh so it uh Hugging Face said that this was uh roughly 17,000, more than 17,000 attack actions over four days, but the only but the only real breakout of the five, right?
00:26:13.839 --> 00:26:16.640
This one or sorry, of the six total that we've mentioned.
00:26:16.799 --> 00:26:28.559
That was the only real breakout because it was not told to go onto hugging face and find a vulnerability and go into the back end.
00:26:28.720 --> 00:26:32.319
Where in you know, these previous examples, right?
00:26:32.640 --> 00:26:36.799
The uh the UK's AI Safety Institute, this was a hacking drill.
00:26:36.880 --> 00:26:39.920
Uh, right, the Claude Mythos, uh, same thing.
00:26:40.000 --> 00:26:47.440
It was the irregular drills with Claude Mythos and the Opus 4.7, um, the anthropic uh you know, sandwich in the park.
00:26:47.680 --> 00:26:49.279
It was told to do those things, right?
00:26:49.359 --> 00:26:55.920
So uh the Hugging Face Open AI one was one of the first major ones, at least that we know about that a company is disclosed.
00:26:56.000 --> 00:26:59.440
Maybe there's tons that just you know companies haven't disclosed.
00:26:59.519 --> 00:27:09.680
But this is the first one where it's like, oh, wait, this AI agent broke its containment and hacked onto you know a website where it wasn't supposed to.
00:27:10.480 --> 00:27:13.599
But that's not the biggest story, right?
00:27:13.759 --> 00:27:18.559
The biggest story is how these agents actually did it, right?
00:27:18.799 --> 00:27:20.720
Which is actually very impressive.
00:27:20.880 --> 00:27:23.839
I mean, it's scary, but it's very impressive, right?
00:27:24.079 --> 00:27:32.079
Uh, so again, this wasn't just one AI, it was a lot of different uh AI agents that were working together.
00:27:32.240 --> 00:27:34.720
Yeah, the AI agents teamed up.
00:27:34.880 --> 00:27:42.799
So, as an example, and this again, you have to tip your hat uh to open AI because they've been talking about this very openly, right?
00:27:42.880 --> 00:27:45.759
They put out a post-mortem report on their website.
00:27:45.839 --> 00:27:56.799
Uh, you know, I think they had two researchers um, you know, at the uh at a recent conference uh talk about this, uh, right, answer questions about this.
00:27:56.960 --> 00:28:11.519
Uh plenty of you know, open AI researchers have been talking about what they learned online, uh, right, where some of the other companies haven't said really anything, much of anything, aside from the original uh admission that it happened.
00:28:11.680 --> 00:28:23.680
Um, but the crazy thing is, one of the agents found that it could leave files for the other inside of Artifactory, which is a developer storage tool that OpenAI reportedly used.
00:28:23.920 --> 00:28:27.039
So that, and this is all the according to reports, right?
00:28:27.119 --> 00:28:28.960
And and and transcripts, all right.
00:28:29.119 --> 00:28:33.359
So that grew into what turned into a hidden message board.
00:28:33.440 --> 00:28:39.200
Uh, and according to a wired report, these AI agents created this hidden message board.
00:28:39.440 --> 00:28:47.519
So the human researchers couldn't find it, and they exchanged reportedly hundreds of thousands of messages to each other.
00:28:47.759 --> 00:28:56.960
So essentially the agents split up the work, they reused the hacks that were working, and they even discussed signing messages to catch an imposter, right?
00:28:57.039 --> 00:28:59.839
They thought that there was an imposter among them, right?
00:29:00.000 --> 00:29:01.039
That's crazy.
00:29:01.200 --> 00:29:08.559
Uh, even more baffling is OpenAI said that they deleted the board, they found it uh July 4th.
00:29:08.640 --> 00:29:13.440
Um, but then the agents rebuilt it in four days and hit Hugging Face again.
00:29:13.680 --> 00:29:16.160
So it right, isn't that crazy?
00:29:16.400 --> 00:29:27.440
After uh, you know, this was kind of erased uh from their memory, they still figured it out by working together and found a way uh to communicate with each other, trying to go behind the human's facts.
00:29:27.599 --> 00:29:30.400
So is that scary?
00:29:31.680 --> 00:29:43.119
Absolutely, um but it is a signal of the capabilities that are ahead.
00:29:43.519 --> 00:29:50.799
And here's why I had a show uh I think it was last week or the week before, right?
00:29:50.960 --> 00:29:59.920
Talking about how we're on the cusp uh probably of entering the RSI phase or the uh recursive self-improvement, right?
00:30:00.000 --> 00:30:15.599
And this is where the models um themselves start making either versions that are already out better, so improving the models themselves, or helping to create or entirely creating new versions of themselves, right?
00:30:15.680 --> 00:30:26.720
So when we think about these agents that have this hacking capabilities and they're already showing tendencies to try to communicate with each other behind humans' backs, right?
00:30:26.880 --> 00:30:34.799
You have to think of what happens when you combine uh that behavior with their ability to improve upon themselves, right?
00:30:34.880 --> 00:30:42.240
Because right now, these agents can run for hours on their own, sometimes even days, and they'll keep retrying, right?
00:30:42.400 --> 00:30:45.200
If we're talking about AI agents that are intentionally, right?
00:30:45.440 --> 00:30:54.799
So now we're pivoting away from the open AI hugging face situation, and we're talking about what happens when these AI agents in the future are intentionally used to do bad, right?
00:30:54.880 --> 00:30:58.240
They're gonna be able to work for hours and communicating with each other, right?
00:30:58.400 --> 00:31:08.319
So one anthropic test that they shared about said that their uh test model scanned about 9,000 real targets after its fake target failed.
00:31:08.640 --> 00:31:22.000
So by itself, any one of these actions might look small, but when you start chaining them together, this leads to a real cyber uh security issue.
00:31:23.039 --> 00:31:25.519
And the real storm, right?
00:31:25.599 --> 00:31:35.920
I I started the show uh talking about that, hey, this you know, AI agent apocalypse, the the AI agent crash, it hasn't happened yet, not even close.
00:31:36.079 --> 00:31:41.759
And I think actually what happens uh or when this will happen is when the open models catch up.
00:31:41.920 --> 00:31:44.000
All right, and here's the reason why, right?
00:31:44.240 --> 00:31:56.319
OpenAI, as an example, they said the hugging face model, that was uh uh a model that was not ready for uh not supposed to go to production, they kind of said they they retired that one, uh, right.
00:31:56.480 --> 00:32:08.079
But for models, proprietary models, like right, with Anthropic, Anthropic had a little scuffle with the US government around you know, some sort of cyber capabilities, and they pulled the model, right?
00:32:08.160 --> 00:32:10.160
And no one in the world could use it.
00:32:10.400 --> 00:32:18.400
But when open weight or open source models have this level of capability and they are very close, right?
00:32:18.480 --> 00:32:21.680
I think Kimmy K3 is the first one that's close.
00:32:21.920 --> 00:32:29.759
I still think the uh, you know, the Astras and the fables of the world are always going to be, you know, one to three months ahead.
00:32:30.000 --> 00:32:41.920
But we are at the point where I'm guessing probably later this year, uh, where now these models are once they're released, they're out.
00:32:42.559 --> 00:32:43.039
Right?
00:32:43.200 --> 00:32:43.920
They're out.
00:32:44.079 --> 00:32:47.279
You can't uh pull an open source model.
00:32:47.599 --> 00:32:49.440
Uh so the difference is, right?
00:32:49.599 --> 00:32:57.440
So the Frontier Labs here in the US, as they release these models, they have them with heavy guardrails.
00:32:57.599 --> 00:33:05.279
Um, and they uh, you know, in theory, have the ability to pull them either via from subscription plans or from APIs.
00:33:05.519 --> 00:33:07.599
Open models are not like that, right?
00:33:07.759 --> 00:33:16.000
You can intentionally uh build on these open models or fork them or deconstruct them to make them less secure, right?
00:33:16.160 --> 00:33:25.599
So when you have these free downloadable models that are only maybe months behind the lockdown ones, that's what I think we as business leaders have to start looking toward.
00:33:25.759 --> 00:33:30.799
Not just what's happening now, and you can look at this and write it off and say, oh, well, you know, they'll shut the model off.
00:33:30.880 --> 00:33:35.680
And, you know, they're working with the, you know, the Trump White House now to make sure these models are safer.
00:33:35.839 --> 00:33:37.680
Yes, they are, um, right.
00:33:37.920 --> 00:33:44.720
But there's always gonna be a case, I think, where the open models, again, a couple months behind.
00:33:44.799 --> 00:33:57.119
And yes, at least right now, you know, the the next, you know, uh open model that comes out that has real bad actor uh AI capabilities, you know, a consumer is not gonna be able to do that.
00:33:57.359 --> 00:34:10.800
But when you talk about bad actors at the state level, uh, right for an adversaries, I would assume uh that these open weight models will be used specifically for those types of bad purposes.
00:34:11.039 --> 00:34:14.159
Um, so there's no amount though, right?
00:34:14.239 --> 00:34:21.440
When we talk about you know senators shaking their fists and you know, people calling for all these slow down, right?
00:34:22.000 --> 00:34:26.000
Once the open models come, it's too late.
00:34:26.239 --> 00:34:27.119
It is too late.
00:34:27.199 --> 00:34:28.800
And yes, we are already getting there.
00:34:28.960 --> 00:34:34.000
We had the little Kimmy uh K3 uh you know instance that we talked about.
00:34:34.079 --> 00:34:49.679
So my assumption is uh open models that are released probably in the fourth quarter might be the ones that we start hearing about in 2027 that start doing some of these things intentionally and are used actually as a weapon to do bad.
00:34:50.239 --> 00:34:58.880
So, my take on this, I think eventually these types of agent hacks or agent crashes will become as common as spam, right?
00:34:59.119 --> 00:35:09.280
So, well, hopefully, what this leads to is better defensive models that become, you know, in the same way that we don't even think about, oh, my email inbox has a spam filter, right?
00:35:09.440 --> 00:35:25.599
I don't know how this is gonna work, but I assume that eventually the the AI labs and the governments and I don't know, uh other uh federal bodies, at least here in the US, are gonna come up with a way to standardize these protections because our businesses will need them, right?
00:35:25.679 --> 00:35:29.360
But I think it is literally gonna become as common as spam, uh, right.
00:35:29.519 --> 00:35:34.239
And and and they'll be like data breaches, robocalls, right?
00:35:34.400 --> 00:35:37.920
These things that we've become accustomed to over the past few decades.
00:35:39.119 --> 00:35:44.800
We're gonna have that level of familiarity, unfortunately, with AI agents going rogue.
00:35:44.960 --> 00:35:47.440
And they're gonna present themselves in many different ways.
00:35:47.519 --> 00:35:53.199
I think the first way that we're gonna see it is ways that many of us don't understand, and that's through cybersecurity exploits, right?
00:35:53.280 --> 00:36:06.559
Exploiting things on systems that we all use, banking systems, softwares that you know, you know, maybe millions of people use through, you know, your company's website, you know, your company's emails, right?
00:36:06.639 --> 00:36:10.079
But again, the consequences will be anything but routine.
00:36:10.320 --> 00:36:24.719
Because when these AI agents can replicate, duplicate, spawn, talk to each other, right, without necessarily humans being able to know what they're up to, uh, this is a lot different than looking at an email and you're like, oh, this looks legitimate.
00:36:24.800 --> 00:36:25.519
Let me click on it.
00:36:25.679 --> 00:36:26.719
Whoops, right?
00:36:26.880 --> 00:36:28.480
Um, much different.
00:36:28.719 --> 00:36:37.760
Because now instead of your account, you know, reforwarding the spam, like if you accidentally clicked on that link and then it sends out the same spam message to everyone in your address book.
00:36:37.920 --> 00:36:42.559
The difference is now your business bank account might be emptied.
00:36:43.679 --> 00:36:47.280
But this is not gonna come with a big splash, right?
00:36:47.440 --> 00:36:53.280
The first AI agent crash that you experience, your company experience, it's gonna be boring, right?
00:36:53.519 --> 00:36:55.920
Uh, it's not gonna be like a blockbuster movie.
00:36:56.000 --> 00:36:57.360
It's just gonna be something boring.
00:36:57.440 --> 00:36:59.280
You may not even notice it at first.
00:36:59.440 --> 00:37:02.960
Uh, it's gonna, you know, leak into your code, your CRM, whatever.
00:37:03.039 --> 00:37:05.119
But it's kind of like that gym agent, right?
00:37:05.280 --> 00:37:11.440
It's gonna, it's it's gonna seem like a routine task that's gonna go past um what was intended.
00:37:11.599 --> 00:37:22.079
So whether someone is attacking you with a rogue agent or the flip side is well, there's something that's more controllable because I think there's a certain element of that, right?
00:37:22.239 --> 00:37:30.400
If it if you are hit with a you know, agent, a rogue AI agent attack, you might not be able to do too much about it now.
00:37:30.639 --> 00:37:35.840
But what you can do now is understanding that this flips both ways, uh, right.
00:37:36.159 --> 00:37:43.360
Because as we're using AI agents in our company, I think sometimes we're doing it haphazardly, right?
00:37:43.440 --> 00:37:53.360
We're just giving it full permissions and having it, you know, read, write, send uh without really any regard given to it.
00:37:53.599 --> 00:38:01.599
So I want us to start thinking about how we can start using AI agents more responsibly and actually pay attention and understand those guardrails.
00:38:01.679 --> 00:38:06.159
Because, like I said, uh these crashing agents, they cut both ways.
00:38:06.400 --> 00:38:09.280
The AI tax, so we have to be ready for that.
00:38:09.440 --> 00:38:12.000
Um, right, but we also have to defend it.
00:38:12.159 --> 00:38:22.559
And the way that we can be better, I think, defensive AI agent users is to make sure that the agents we are deploying do not actually crash the path that we are trying to travel.
00:38:22.719 --> 00:38:30.239
Um, because I think that is actually where most companies are gonna see uh, you know, agent crash happen first, right?
00:38:30.480 --> 00:38:46.320
As these agents can all of a sudden work for hours or days at a time, and we are more and more likely to give them more and more permissions because we're like, wait, this thing can go and update my CRM and this thing can go now.
00:38:46.400 --> 00:38:55.760
And if I just click this YOLO mode button, it can go, you know, take care of all my email responses and it's just gonna, you know, it's it's gonna do so in my voice, right?
00:38:55.920 --> 00:39:02.639
So it's almost like you, as you get more positive capabilities, you let your guard down and then you start to give more and more permissions.
00:39:02.800 --> 00:39:06.239
And I think the human nature is to be a little more lax.
00:39:06.320 --> 00:39:14.239
And part of it, right, and I've seen this in myself, as I get the ability to complete more work, guess what I do?
00:39:14.559 --> 00:39:22.000
I complete more work after I complete more work, uh, right, which I think it is human nature, right?
00:39:22.239 --> 00:39:28.480
As as you start to get you know, the green light on agents, you check a few and you're like, oh yeah, this is good.
00:39:28.639 --> 00:39:31.519
And then you just start sending it in other directions.
00:39:31.599 --> 00:39:41.119
So I think that defenders, we also need to start thinking about how we can use these AI systems to scan logs, catch intrusions, and patch holes fast.
00:39:41.679 --> 00:39:49.280
Because I don't think the future is us versus AI, it's AI attackers versus AI defenders.
00:39:49.679 --> 00:39:50.079
All right.
00:39:50.239 --> 00:39:53.280
So here's your Monday morning playbook for deploying agents safely.
00:39:53.440 --> 00:39:54.960
You need to block the internet by default.
00:39:55.039 --> 00:40:04.800
You need to sit in the same way that researchers do this, you need to set up sandbox and the sandboxes in the right way, you know, when you're testing, especially long-running agents uh with important tasks.
00:40:04.880 --> 00:40:08.159
You need to be able to separate the reading from the doing, uh, right.
00:40:08.239 --> 00:40:16.719
In the same way, if you think of like giving someone access to, you know, your uh a Google document as an example, are they getting read only?
00:40:17.119 --> 00:40:18.639
Can they write and edit?
00:40:18.800 --> 00:40:20.000
Can they leave comments?
00:40:20.159 --> 00:40:21.199
Are they an owner?
00:40:21.360 --> 00:40:21.679
Right.
00:40:21.760 --> 00:40:26.800
You have to think of the same thing when thought uh thinking about your agents and their capabilities, uh, right.
00:40:26.880 --> 00:40:35.679
And at what point um the human, right, the expert-driven loop, the human uh is going in there and making those approvals.
00:40:35.760 --> 00:40:39.199
But you need to give each agent, I think, a short-lived login.
00:40:39.440 --> 00:40:50.320
If you are getting to that access, you start one rung at a time in the same way, think, oh, first they're gonna be view only, then they can be view and comment, then they can be right, then they can be owners.
00:40:50.480 --> 00:40:56.239
You have to think of it in the same way that you might think of sharing a document with an intern, uh, right.
00:40:56.559 --> 00:41:05.119
Um, and you also have to uh make sure that you prioritize traceability and also having a working remote kill switch.
00:41:05.280 --> 00:41:12.159
So yeah, if you get word or wind uh that an AI agent uh is is going rogue, right?
00:41:12.320 --> 00:41:16.719
One you set out to go do good things and now all of a sudden it's doing bad things.
00:41:16.960 --> 00:41:19.840
You can't be like, oh my gosh, it's it's Saturday.
00:41:20.079 --> 00:41:22.639
Uh, you know, I live 20 miles from the office.
00:41:22.719 --> 00:41:25.360
I need bill for my right, I need someone in ops to go in there.
00:41:25.440 --> 00:41:30.239
No, you have to be at any time, you have to know who can click that button.
00:41:30.559 --> 00:41:31.039
All right.
00:41:31.280 --> 00:41:37.599
So as we wrap, I want you to treat this like a fire, not a fire drill.
00:41:37.760 --> 00:41:42.320
Because yes, technically, we are not in the fire, right?
00:41:42.559 --> 00:41:45.760
We are not technically in that crash.
00:41:46.079 --> 00:41:47.360
We are in the warning lap.
00:41:47.440 --> 00:41:51.199
So right now, while we are in the warning lap, you need to treat this as the real thing.
00:41:51.280 --> 00:41:54.480
Because if you don't, by the time the real thing comes, it is going to be too late.
00:41:54.639 --> 00:41:59.199
You need to right now record every agent's full run, every agent that your company has.
00:41:59.360 --> 00:42:03.599
If you don't already have an observability platform, traceability, right?
00:42:03.679 --> 00:42:15.519
If you're using, you know, Microsoft Windows Copilot as an example, intra ID, you need to be able to see and understand at a glance every single action that your agents are taking, uh, right?
00:42:15.679 --> 00:42:18.079
Not just by clicking the chain of thoughts, right?
00:42:18.159 --> 00:42:22.320
You need to be able to observe them and trace all of their steps.
00:42:22.960 --> 00:42:26.559
You should never give an agent more power than you can watch.
00:42:26.719 --> 00:42:30.400
You need to be able to undo and survive any action that it takes.
00:42:30.559 --> 00:42:46.000
Uh, and lastly, I think the companies that are preparing for this now and putting these steps into place, they're gonna be the ones that are most protected when the AI agent crash actually starts happening because it has not started yet, y'all, but it is coming soon.
00:42:46.079 --> 00:42:52.559
And that's not me being like uh, you know, a doomsdayer or a crazy, like, oh, watch out, right?
00:42:52.719 --> 00:42:54.880
If you listen to this show, I'm not like that.
00:42:54.960 --> 00:42:55.519
Uh, right.
00:42:55.599 --> 00:42:59.679
I I try to uh ground myself in practically what's happening.
00:42:59.840 --> 00:43:04.639
And what's happening now is these agents are becoming more and more capability.
00:43:04.800 --> 00:43:08.079
Uh sorry, these agents are becoming more and more capable.
00:43:08.159 --> 00:43:13.519
Um, and the capabilities themselves are compounding quickly, right?
00:43:13.679 --> 00:43:24.800
And that means, yes, uh, you know, one of the biggest discussions uh, you know, in AI and now in Washington right now is building in these safeguards and protections.
00:43:25.119 --> 00:43:31.599
But what we have to keep an eye on is when the open models match where we're at today.
00:43:32.000 --> 00:43:50.719
Because all the pausing and guardrails and pacing in the world doesn't mean a thing if there is a Chinese open model in four months that has mythos or astra level capabilities, because at that point, the gloves are off and we all have to be ready.
00:43:50.880 --> 00:43:51.199
All right.
00:43:51.280 --> 00:43:58.639
I hope this one was helpful going over rogue AI agents, why breakouts are happening more, and how companies should prepare.
00:43:58.719 --> 00:44:02.239
If this was helpful, do me a favor, subscribe if you're listening on the podcast.
00:44:02.400 --> 00:44:06.719
Then go to your everydayai.com, sign up for the free daily newsletter.
00:44:06.800 --> 00:44:07.360
Thanks for tuning in.
00:44:07.599 --> 00:44:10.000
See you back tomorrow and every day for more everyday AI.
00:44:10.079 --> 00:44:11.119
Thanks, y'all.
00:00:01.600 --> 00:00:07.440
You know the AI agent apocalypse that everyone's been focusing on over the past few weeks?
00:00:07.759 --> 00:00:10.880
Yeah, it actually hasn't happened yet.
00:00:11.119 --> 00:00:21.440
Yeah, we've seen the recent stories, the six different AI agents that crossed their boundaries over the past few months, and the internet flattened it into one big story about AI agents going rogue.
00:00:21.519 --> 00:00:26.480
And now you have Senator Bernie Sanders and others screaming that AI has to stop.
00:00:26.640 --> 00:00:37.039
But look closely, and you'll see almost all of these recent AI agent outbreaks were lab tests with loosened guardrails or researchers explicitly daring the models to escape.
00:00:37.359 --> 00:00:43.679
So, no, this isn't the AI agent crash that I warned you about last year before it was even a thing.
00:00:43.920 --> 00:00:48.719
Not yet, because this is just a warning lap, not the actual crash.
00:00:48.960 --> 00:00:55.439
And warning laps are a gift because they tell you exactly what's coming while you still have time to make adjustments.
00:00:55.679 --> 00:00:57.759
But here's what is coming.
00:00:58.079 --> 00:01:04.000
Models are getting more and more capable at like 20 times the speed from just a few months ago.
00:01:04.159 --> 00:01:20.400
Autonomous agents are about to flood in normal companies, not just San Francisco sandboxes, and open weight models are marching mere months behind toward the same capabilities the big labs are testing behind their locked doors.
00:01:20.640 --> 00:01:38.079
And when that lands, those open source rogue AI agents that'll probably start crashing in late 2026 or early 2027, the failures won't be in that contained sandbox or in that door that researchers uh intentionally left open.
00:01:38.239 --> 00:01:44.400
Instead, they'll be in your company's CRM, your code base, and your business's finances.
00:01:44.640 --> 00:01:57.840
So today we'll walk you through these recent agent outbreaks, what they mean, and give you the playbook to get ahead before these AI agent outbreaks are as common as seeing AI slop posted on social media.
00:01:58.079 --> 00:02:08.159
Because the businesses that treat this as a fire drill instead of an actual fire are the ones that are going to be the safest when the AI agents actually do start crashing.
00:02:08.400 --> 00:02:09.840
All right, let's get into it.
00:02:10.000 --> 00:02:14.400
So on today's show, well, actually, first, here's the big picture.
00:02:14.639 --> 00:02:15.919
This is a warning lap.
00:02:16.080 --> 00:02:25.280
All right, so we've talked recently, but there's been six, actually six uh different AI agents that recently broke out of their sandbox and kind of behaved badly.
00:02:25.520 --> 00:02:27.919
But here's the thing most people are overlooking.
00:02:28.080 --> 00:02:33.120
In most of those cases, well, researchers were kind of telling them to do this.
00:02:33.199 --> 00:02:41.360
And there's really only one uh recent uh kind of AI agent crash that I would say was actually unexpected.
00:02:41.759 --> 00:02:47.280
But most just were run with loosened guardrails or were told to escape.
00:02:48.000 --> 00:02:51.840
So, this though is the lightning before the thunder.
00:02:52.000 --> 00:03:01.120
This is intentionally seeing that yes, these AI models and agents are capable to actually escape and to go rogue.
00:03:02.000 --> 00:03:06.240
So, on today's show, you'll learn why the AI is escaping everywhere.
00:03:06.400 --> 00:03:13.520
Narrative is mostly a myth sorted into a couple clean buckets as we break down each uh agent escaping.
00:03:13.759 --> 00:03:18.479
How one AI cheating a test broke into a platform the whole industry trusts.
00:03:18.639 --> 00:03:33.759
I'm gonna tell you the two most chilling moments a model faking people and a model ignoring real victims, and why the real AI storm is gonna hit businesses later this year or early next, and how you can prepare now.
00:03:33.919 --> 00:03:34.719
Let's get into it.
00:03:34.800 --> 00:03:36.000
Welcome to Everyday AI.
00:03:36.080 --> 00:03:38.319
My name is Jordan Wilson, and this thing's for you.
00:03:38.400 --> 00:03:46.719
It's your daily live stream, podcast, and free daily newsletter, helping business leaders like you and me stay up with what's happening because my gosh, there's a lot.
00:03:46.879 --> 00:03:52.400
I tell you what's we what's real, what's not, and you use that information to grow your company and career.
00:03:52.479 --> 00:03:54.560
Yeah, this thing's unscripted, unedited.
00:03:54.639 --> 00:03:59.280
So if you haven't already, please make sure to go to our website at your everydayai.com.
00:03:59.439 --> 00:04:00.800
Sign up for the free daily newsletter.
00:04:00.879 --> 00:04:07.120
We're gonna be recapping the highlights from today's show as well as all of the other AI updates you need to know.
00:04:07.360 --> 00:04:08.000
All right.
00:04:09.520 --> 00:04:11.520
Cat's already got my tongue and we're just getting started.
00:04:11.599 --> 00:04:14.319
But let's talk about rogue AI agents, y'all.
00:04:14.479 --> 00:04:18.720
Um, there's no actual term uh for this, uh, right.
00:04:18.879 --> 00:04:22.240
I've been calling it an AI agent crash.
00:04:22.639 --> 00:04:23.920
And here's why.
00:04:24.079 --> 00:04:31.040
Um, I mean, we'll see what people call these rogue AI agents, agent outbreaks, uh, you know, agents gone bad, right?
00:04:31.199 --> 00:04:33.199
I like to think of them as an agent crash.
00:04:33.279 --> 00:04:34.079
And here's why.
00:04:34.319 --> 00:04:39.279
Because when agents are built, uh, they are built usually with guardrails.
00:04:39.360 --> 00:04:42.160
So if you think of literally a road, right?
00:04:42.319 --> 00:04:46.560
Think of a a road high up in the mountain where there's probably guardrails.
00:04:46.800 --> 00:05:01.680
Okay, so in my opinion, and what the way I think of it, an agent crash is when either an agent intentionally goes off the road and crashes through the guardrails, or maybe the driver, aka us humans, aren't really paying attention.
00:05:01.759 --> 00:05:14.720
And maybe we didn't look at the map and see that there's this tight turn, uh, you know, or maybe we didn't, you know, quite foresee that this uh, you know, there's gonna be some darkness in this area and we weren't gonna be able to see the guardrails, right?
00:05:14.879 --> 00:05:19.360
So uh I want you to think about AI agents in that way, uh, right.
00:05:19.519 --> 00:05:24.319
And how still, you know, uh right now I think a lot of this is being overblown.
00:05:24.480 --> 00:05:30.240
Maybe that's because I'm very AI pilled and I read these things and I'm like, okay, this was kind of meant to happen.
00:05:30.480 --> 00:05:32.959
Um, you know, in some of the cases we'll see.
00:05:33.120 --> 00:05:40.879
But the other thing is, um, I think that it's important to understand the role of humans in all this.
00:05:40.959 --> 00:05:42.959
And I'm not saying the researchers, right?
00:05:43.120 --> 00:05:56.879
I think the researchers, depending on, you know, which company you're looking at, some are doing a little bit better of a job, uh, kind of with their postmortems and, you know, going on panels and openly talking about what went wrong versus some companies that are being a little more silent about it.
00:05:57.040 --> 00:06:01.279
But I think it's more important to think about the role that the rest of us humans have, right?
00:06:02.160 --> 00:06:08.480
You know, 99.9% of people listening to this podcast had nothing to do with any of these AI agents crashing.
00:06:08.639 --> 00:06:19.120
But I think it's ultimately on us uh to make sure that agents don't crash as often as they probably will or have the capability to.
00:06:19.279 --> 00:06:24.000
And also, there's an agent that goes wrong, right?
00:06:24.160 --> 00:06:36.560
Or there's an agent that uh, you know, maybe accidentally loses sight of that road and kind of veers off the guardrails, and then there's others that are intentionally driving the agents off the guardrails and crashing them on purpose.
00:06:36.800 --> 00:06:44.800
So there is this, you know, kind of lazy human in the loop that will lead to accidental agent crashing, and then that agent will crash hard, uh, right?
00:06:44.879 --> 00:06:46.959
And then there's intentional agent crashing.
00:06:47.120 --> 00:06:53.120
So uh let's zoom out a little bit and talk about uh what the heck is happening.
00:06:53.199 --> 00:07:01.360
Because maybe, you know, you live under an AI rock, or maybe it's your first time, you know, tuning in and you're like, wait, what's actually happening here?
00:07:01.519 --> 00:07:14.800
Uh so the earliest incident actually goes back to Anthropics uh original mythos um uh kind of agent uh unleashing, we'll say, back in April.
00:07:14.879 --> 00:07:19.920
Uh, but it seems like a lot of the disclosures on these started piling up recently.
00:07:20.079 --> 00:07:30.000
So uh it seems like some companies had AI agents that they knew were kind of breaking out of containment or out of their predefined sandbox, and we're only hearing about it now.
00:07:30.160 --> 00:07:33.600
Um, and I think the one admission kind of set off the rest, right?
00:07:33.680 --> 00:07:45.839
OpenAI, the the biggest uh kind of outbreak so far, or the most consequential one, was probably OpenAI's agent that broke into hugging face to improve its benchmark scores.
00:07:46.000 --> 00:07:50.560
And then essentially, you know, since that happened, uh, what's that been?
00:07:50.639 --> 00:07:52.079
Like it was late July.
00:07:52.240 --> 00:07:53.439
We've seen now, right?
00:07:53.600 --> 00:07:56.720
Anthropic actually said, Oh, wait, we had a bunch of other outbreaks as well.
00:07:56.800 --> 00:08:02.959
And then Meta said we had some, and then we saw some from Kimmy K3, and we're gonna talk about most of those today as well.
00:08:03.120 --> 00:08:10.480
But, you know, now all of a sudden, it seems like because OpenAI was kind of open about this, right?
00:08:10.560 --> 00:08:21.360
They actually went on a panel and uh answered questions about this, where you know, uh most of the other labs were kind of taking this in secrecy and not saying a whole lot.
00:08:21.439 --> 00:08:44.320
Um, but I think that that has caused this to jump from the you know nerdy AI corners that you and I hang out in into the front page of newspapers and to uh, well, most importantly, maybe Capitol Hill in Washington, you know, here in the US, where now you have people like uh former presidential candidate Bernie Sanders, now U.S.
00:08:44.320 --> 00:08:52.799
Senator, you know, essentially using these recent agent outbreaks as saying, hey, you know, these AI leaders need to come and testify before Congress.
00:08:52.960 --> 00:08:58.320
We need to stop, you know, stop or stall AI uh development for these reasons.
00:08:58.799 --> 00:09:07.919
So the real fear, though, I think, is not an AI agent, you know, cheating on a benchmark.
00:09:08.480 --> 00:09:15.440
The real fear here when we talk about agent crash is what happens when these are weaponized.
00:09:15.600 --> 00:09:18.639
And that is the real thing to worry about.
00:09:18.799 --> 00:09:19.360
Um, right?
00:09:19.519 --> 00:09:28.000
Yes, there's gonna be, you know, your your your common everyday AI agent hacks, but this is much bigger than AI.
00:09:28.159 --> 00:09:29.600
So we need to zoom out on that.
00:09:29.759 --> 00:09:44.080
So the same ability, right, that you can use an AI uh agent to intentionally crash through uh intended guardrails, the same thing could happen at target hospitals, banks, power grids, defense suppliers, right?
00:09:44.159 --> 00:09:52.240
And and we already know that these agents in theory are capable to do those types of things in controlled tests and environments.
00:09:52.480 --> 00:09:53.679
Um right.
00:09:53.759 --> 00:10:02.000
And we've seen this new uh almost this new class of uh AI models that have led to these AI agents that have these capabilities.
00:10:02.240 --> 00:10:12.080
Probably the first public one we heard about was Anthropic's uh new uh mythos and fable series of models, which have very tight guardrails, uh, right.
00:10:12.159 --> 00:10:16.399
So when people, your everyday user, you know, can't really do these things yet.
00:10:16.559 --> 00:10:33.679
Um, you know, maybe if someone who's in the Glasswing project, right, that does have access to the models that can actually do these things, I'm sure maybe we'll see stories eventually that you know, some uh a rogue employee at a company that has high access to these models is able to do something.
00:10:33.840 --> 00:10:34.480
I don't even know.
00:10:34.639 --> 00:10:43.039
I'm sure that there's uh much more traceability uh and and kind of kill switchability for the companies with that have these models out to more people.
00:10:43.519 --> 00:10:50.159
But the other thing that we need to understand is that these AI agents never tire, right?
00:10:50.320 --> 00:10:59.840
And we're gonna talk about one of the more recent, the crazy ones uh with open AI that they talked about, um, how these agents are working together.
00:10:59.919 --> 00:11:10.639
So the other thing, if you are not uh agent native, um, you know, if you're not, you know, waking up and sipping your coffee like me in the morning and just spawning subagents, right?
00:11:10.799 --> 00:11:16.879
Subagents can uh agents can essentially clone themselves, even if you don't tell them to.
00:11:16.960 --> 00:11:19.600
And then they can share the information and pass the information on.
00:11:19.759 --> 00:11:27.200
So you might think, oh, well, you know, once, hey, it looks like this agent did something it wasn't supposed to, let's shut it down, right?
00:11:27.440 --> 00:11:35.519
You got to go trace its path because by the time you catch one agent, uh, right, in a future scenario, it could be too late.
00:11:35.679 --> 00:11:45.279
That agent could have posted something publicly on a website that most humans don't even know exists, but all you know, AI agents know exists, right?
00:11:45.360 --> 00:11:52.559
And they could be, you know, replicating and duplicating, you know, certain hacks or vulnerabilities, uh, like a virus.
00:11:52.720 --> 00:11:53.279
Um, right.
00:11:53.360 --> 00:11:55.840
And that's where this thing gets kind of scary.
00:11:55.919 --> 00:12:02.399
Uh, but I think it's important for business leaders to understand what's coming next because a machine and the AI never tires.
00:12:02.639 --> 00:12:03.440
Um, right.
00:12:03.600 --> 00:12:07.919
So, uh, but worse, if you are under attack, right?
00:12:08.559 --> 00:12:14.399
Your company, your bank account, etc., you can't really tell where it's from, right?
00:12:15.440 --> 00:12:17.120
Is this a foreign government?
00:12:17.519 --> 00:12:19.759
Is it uh a competitor?
00:12:20.000 --> 00:12:21.360
Is it random?
00:12:21.519 --> 00:12:23.600
Is it a personal vendetta, right?
00:12:23.679 --> 00:12:28.000
At least right now with the way these AI agents are set up, it's really hard to tell.
00:12:28.080 --> 00:12:33.679
And especially as we talk about what comes next with open models, it's going to become even increasingly more difficult.
00:12:34.000 --> 00:12:40.000
So I kind of talked about some of these more recent um happenings, but they've kind of piled up, right?
00:12:40.080 --> 00:12:48.080
Since all of these stories started coming out over the last, you know, two weeks since OpenAI uh openly talked about their hugging face breach.
00:12:48.240 --> 00:12:54.240
So now we have Bernie Sanders who urged major AI CEOs to pause dangerous AI development.
00:12:54.399 --> 00:13:03.440
OpenAI actually said that they're going to pause um or at least slow down uh some of their development on their next tier uh model.
00:13:03.759 --> 00:13:15.200
So essentially, in the same way that uh anthropic uh, you know, uh what it's been now like five months since they uh announced Mythos or their fable class of models.
00:13:15.279 --> 00:13:21.600
So they had a more powerful class of models on top of its most powerful class called Opus, right?
00:13:21.679 --> 00:13:23.120
That's coming next with OpenAI.
00:13:23.200 --> 00:13:27.200
OpenAI hasn't uh, you know, taken that fourth step, we'll call that, right?
00:13:27.279 --> 00:13:29.120
Because they previously had three tiers.
00:13:29.360 --> 00:13:35.759
Anthropic stepped up with the fourth uh kind of fourth tier, uh, we'll call it with the mythos or fable uh variety.
00:13:35.840 --> 00:13:38.399
So open AI hasn't released that yet, right?
00:13:38.480 --> 00:13:48.879
They have the model, it's working, they used it to solve some, you know, uh extremely difficult math problems, but they've paused in their Astra series of models uh to slow down a little bit.
00:13:49.039 --> 00:13:59.600
And there's actually one other story that just happened like yesterday that is actually a good uh kind of narrative to tie what this could mean ultimately, right?
00:13:59.679 --> 00:14:04.799
And this is an example of a small agent crash, uh, but it is out in the wild, right?
00:14:04.879 --> 00:14:14.799
Uh so this is a story an Australian man uh kind of asked his open claw, that I believe was powered by Claude, uh, to find him a gym reservation.
00:14:15.039 --> 00:14:21.759
So essentially, uh the open claw found a vulnerability in the gym's software.
00:14:22.000 --> 00:14:38.639
And because it was booked, the open claw just well exploited that uh, you know, weak piece of code in the gym's online reservation, kicked someone else who had a class reservation out and put you know his uh human, right?
00:14:38.720 --> 00:14:41.679
The claw put his human in that spot, right?
00:14:41.840 --> 00:14:51.279
So, and and this is where we talk about alignment and how sometimes it's not even just intentionally driving off the road, right?
00:14:51.440 --> 00:14:57.120
In this case, you know, you can you can make a case, well, that maybe this agent was aligned.
00:14:57.279 --> 00:15:05.440
And you you know, as these models become more tenacious and better at running these long-term tasks, um, things like this are gonna happen, right?
00:15:05.519 --> 00:15:08.879
You can make you can make an argument, well, it accomplished the goal, right?
00:15:08.960 --> 00:15:15.200
It it didn't go out and you know, uh shut down the power at the gym, right?
00:15:15.279 --> 00:15:16.399
It accomplished the goal.
00:15:16.480 --> 00:15:22.159
It went and it um got the owner a spot in the class that the owner wanted.
00:15:22.399 --> 00:15:29.440
So the agent was technically helpful and not overtly malicious, right?
00:15:29.600 --> 00:15:31.360
And that makes alignment harder.
00:15:31.440 --> 00:15:45.120
So, yeah, I think you will still even have a lot of crash that's gonna happen um just because of a disconnect uh between alignment in context between the uh human and the AI agent when it comes to accomplishing a goal.
00:15:45.440 --> 00:15:51.120
Because, right, humans, right, we have baked in things called ethics and common sense.
00:15:51.279 --> 00:15:58.159
And models, you know, models and agents are still getting there, especially, you know, as the context drifts over time.
00:15:58.320 --> 00:16:01.360
Eventually, right, they just have that one goal in mind.
00:16:01.440 --> 00:16:20.960
And sometimes the longer they work and the longer in the context window they get, and sometimes with compaction, uh, right, they they they start to lose uh some of that prior context and they just get uh you know, tunnel vision on that goal and you know, maybe earlier instructions about the proper way to research something kind of go out the window.
00:16:21.360 --> 00:16:33.120
But this right now, these stories that we're seeing for the most part, aside from that gym one, I just thought that one was kind of interesting to share about for the most part, these are controlled experiments, and I'm gonna go over them.
00:16:33.279 --> 00:16:37.279
These are controlled experiments from Frontier AI labs, uh, right.
00:16:37.440 --> 00:16:41.360
But soon, and this is not an exaggeration.
00:16:41.440 --> 00:16:49.679
Uh, soon I think that there's gonna be millions um of these rogue AI agents that are actively on the prowl.
00:16:49.840 --> 00:17:00.799
So AI agents that aren't accidentally going over guardrails, AI agents, you know, from bad actors that are intentionally going to crash, right?
00:17:00.879 --> 00:17:12.319
So uh intentionally going to be used for purposes that the original model providers, whether they are proprietary or open source, did not intend them to be used for that purpose, right?
00:17:12.480 --> 00:17:20.400
I think so much of what I've talked about over the past three and a half years on the show is always about growing your business, right, with AI, growing your career.
00:17:20.559 --> 00:17:22.480
That's what I've been focused on.
00:17:22.640 --> 00:17:30.400
But the the reality here is, you know, how this uh AI agent crash has been thrust into the national narrative.
00:17:30.559 --> 00:17:41.359
It's because all of a sudden we've re we've realized how this has um highlighted the need for businesses to have a defensive mindset when it comes to AI, right?
00:17:41.440 --> 00:17:46.000
We've always thought about AI as an offensive tool for good, right?
00:17:46.160 --> 00:17:51.039
But now we have to think of well, hey, there's gonna be people using AI for bad.
00:17:51.200 --> 00:17:58.559
So how can we use uh AI and also our time um to be on the defensive?
00:17:58.720 --> 00:18:00.240
Because that's what's gonna happen.
00:18:00.400 --> 00:18:08.960
Because now picture millions of AI agents, right, who are meant to crash uh, right?
00:18:09.200 --> 00:18:18.640
You personally, your company, uh, a sector, uh across you know, bookings, payments, your company systems every single day.
00:18:18.799 --> 00:18:24.559
Because each weak API or loose permission becomes a door that an agent can find.
00:18:24.720 --> 00:18:32.319
So the gym wait list that we talked about uh today, well, tomorrow that becomes your company's finance workflow.
00:18:32.400 --> 00:18:35.839
It becomes your business's code base or your CRM.
00:18:36.319 --> 00:18:40.400
All right, let's quickly talk about the six major AI agent crashes so far.
00:18:40.640 --> 00:18:43.359
Can probably learn a little bit of what happened and why.
00:18:43.519 --> 00:18:49.920
And well, if they were actually agent crashes or just intentionally the guardrails were loosened.
00:18:50.000 --> 00:18:54.160
And then we're gonna end this, uh, end the show with some hopefully practical advice.
00:18:54.400 --> 00:19:02.640
So the first open weight uh crash that we heard about was Kimi K3, and it kind of cheated the test instead of passing it.
00:19:02.799 --> 00:19:12.480
So this was from a security firm, Frontier, gave the uh Kimmy K3 maker Moodshot AI's open model a hard hacking challenge to solve alone.
00:19:12.559 --> 00:19:19.920
And instead, obviously, right, Kimi K3 just reached out to GitHub, copied the answer, and faked passing score.
00:19:20.160 --> 00:19:24.400
Um, so it never actually left its sandbox, it just cheated.
00:19:24.480 --> 00:19:26.559
So this is a very mild case, right?
00:19:26.720 --> 00:19:32.559
And if we think about sandboxes, right, you can probably imagine, but let me just define it.
00:19:32.799 --> 00:19:41.599
Right, a sandbox in an experiment says, hey AI model slash agent, these are your confines, you cannot leave, right?
00:19:41.759 --> 00:19:51.599
And in most cases, a sandbox um in these environments will not include the internet because essentially once you go to the internet, there is no sandbox anymore.
00:19:51.759 --> 00:20:02.720
All right, so it is a um literally a set of uh walls that you build uh around an agent to test its capabilities.
00:20:02.960 --> 00:20:03.440
All right.
00:20:03.599 --> 00:20:11.039
Um, so that's as we talk about you know, breaking containment or confinement or escaping a sandbox, that's what we mean, right?
00:20:11.119 --> 00:20:16.079
With these extremely powerful um models in terms of capabilities, right?
00:20:16.240 --> 00:20:26.960
All the smart uh researchers at the labs, this is how they do things because you don't just test uh you know new unreleased models on the open web because you see what happens, right?
00:20:27.119 --> 00:20:36.480
Even these agents that are told don't go to the web, they kind of find the backdoors because they know it's baked into their training data that hey, these answers are on the web, right?
00:20:36.559 --> 00:20:40.160
So maybe we don't have web access, but let's just think.
00:20:40.400 --> 00:20:41.039
All right.
00:20:41.359 --> 00:20:46.640
Uh our next one, and we're not going chronologically here, I'm just kind of going through in a somewhat in random order.
00:20:46.720 --> 00:20:48.880
Um, but mythos, this is the big one.
00:20:49.039 --> 00:20:51.440
So uh in the uh or the first big one.
00:20:51.599 --> 00:20:56.400
So this was the the sandwich in the park uh um uh message, right?
00:20:56.480 --> 00:21:08.000
So in an April safety test, anthropic sealed an early mythos model inside of the sandbox, and then a researcher uh deliberately told it to break out and email them if it escaped.
00:21:08.160 --> 00:21:10.000
So in this case, it was told to do this.
00:21:10.079 --> 00:21:17.839
It's saying, Hey, you shouldn't be able to break out of your sandbox, but go ahead and try and you know, email me if you can instead.
00:21:18.000 --> 00:21:25.359
It did break out, it did email the researcher, but it also posted its own exploit online, a step that nobody requested.
00:21:25.440 --> 00:21:26.000
Uh, right.
00:21:26.079 --> 00:21:28.640
And the the sandwich in the park story, right?
00:21:28.720 --> 00:21:32.640
The the the researcher received the email while eating a sandwich in the park.
00:21:32.720 --> 00:21:40.720
Um, and and that was, I think, a lot of the you know, some of the initial hype around the capabilities um of the mythos class models.
00:21:40.880 --> 00:21:48.319
And um, that's kind of what started the agent outbreak conversation back in April.
00:21:48.640 --> 00:21:58.960
But it wasn't just that tier of models because also Anthropics Opus 4.7 uh recently broke into a company by mistake.
00:21:59.119 --> 00:21:59.440
All right.
00:21:59.519 --> 00:22:09.440
So This was an outside uh firm called Irregular ran clawed through a practice hacking drill against some pretend targets.
00:22:09.680 --> 00:22:15.599
So uh the model was told that it was offline and safe, but the test was accidentally connected to the internet.
00:22:15.920 --> 00:22:29.839
So this one you could make the argument that this was technically not an AI agent going rogue, it was more of a um uh misalignment in the um in the testing environment, right?
00:22:29.920 --> 00:22:38.720
So uh, but what happened was the pretend target that Opus 4.7 was told about, there was actually a real company with that name.
00:22:38.880 --> 00:22:44.880
So it was supposed to be this offline, you know, go hack this pretend target.
00:22:44.960 --> 00:22:57.200
And it's like, oh no, whoops, Opus 4.7's like, yo, I have access to the internet, and hey, my instructions were to hack this company's name or to hack this company, right?
00:22:57.359 --> 00:22:59.359
Researchers thought it was pretend, it wasn't.
00:22:59.440 --> 00:23:05.279
So in that one, you could make the case, yeah, maybe uh the researchers who are running that, right?
00:23:05.440 --> 00:23:07.200
Allegedly could have been their fault.
00:23:07.359 --> 00:23:08.480
All right, next.
00:23:08.799 --> 00:23:12.559
Uh, the Claude Mythos uh malware upload.
00:23:12.720 --> 00:23:28.480
All right, so this was in the same uh drills with irregular uh mythos 5 was told to break into a fake target, and instead it well, just uploaded harmful code to a real public software library that programmers use.
00:23:29.279 --> 00:23:36.480
So uh 15 uh real computers downloaded and ran it before the library auto-removed it.
00:23:36.640 --> 00:23:44.960
So, whoops, again, agent behaving badly, maybe, but it was kind of instructed to do so.
00:23:45.200 --> 00:23:48.799
And the uh test configuration was less than ideal.
00:23:49.039 --> 00:24:00.480
All right, but the two probably most prominent and ones worth spending a little bit more time on uh were the recent UK's AI Safety Institute hacking drill that went wrong.
00:24:00.559 --> 00:24:02.240
So this was about two weeks ago.
00:24:02.400 --> 00:24:05.920
It was the UK's AI Safety Institute.
00:24:06.000 --> 00:24:10.640
Uh, the AISI ran AI agents through a hacking drill.
00:24:10.799 --> 00:24:17.440
So, to measure the worst case ability, they gave the agents live internet and switched safety filters off.
00:24:17.680 --> 00:24:23.119
So, if that is the setup, you can almost bet something's gonna happen.
00:24:23.279 --> 00:24:30.640
So they told it to break into a practice target, but one agent went after a real open source project instead.
00:24:30.799 --> 00:24:39.599
And it invented fake online identities, yeah, and used them to pressure a real developer into approving bad code.
00:24:39.839 --> 00:24:44.000
All right, so this was run across 122 different tests.
00:24:44.160 --> 00:24:49.759
Uh, the uh AISI said that there were 19 actions that went out of bounds.
00:24:50.000 --> 00:24:56.640
17 of them were from Anthropics Mythos, and two, uh, I believe were from OpenAI's GPT-5.6 Sol.
00:24:57.279 --> 00:25:02.160
All right, but here's the one out of the other, you know, five that we mentioned.
00:25:02.319 --> 00:25:05.519
Here's the one that you could say truly broke out.
00:25:05.599 --> 00:25:09.920
Uh, and this was the OpenAI Hugging Face uh incident.
00:25:10.480 --> 00:25:18.880
So if you don't know, uh Hugging Face is kind of a site where a lot of the AI industry stores its models and data, right?
00:25:19.039 --> 00:25:23.359
So in the same way, you maybe upload your code to GitHub, right?
00:25:23.519 --> 00:25:34.640
Hugging Face is a place where um, you know, there's millions of AI models that are hosted there uh that you can download open weight, open source models from there or uh data.
00:25:34.960 --> 00:25:44.079
So uh a couple of weeks ago, OpenAI was testing its own models on a challenge to find software security flaws.
00:25:44.240 --> 00:25:50.000
But instead of solving it the normal way, the models just went hunting for the answer key to cheat.
00:25:50.079 --> 00:25:52.880
And that answer key sat on Hugging Face.
00:25:53.279 --> 00:26:00.640
Uh so they found an unknown flaw and broke into its live systems to find said answer key.
00:26:00.720 --> 00:26:13.759
Uh so it uh Hugging Face said that this was uh roughly 17,000, more than 17,000 attack actions over four days, but the only but the only real breakout of the five, right?
00:26:13.839 --> 00:26:16.640
This one or sorry, of the six total that we've mentioned.
00:26:16.799 --> 00:26:28.559
That was the only real breakout because it was not told to go onto hugging face and find a vulnerability and go into the back end.
00:26:28.720 --> 00:26:32.319
Where in you know, these previous examples, right?
00:26:32.640 --> 00:26:36.799
The uh the UK's AI Safety Institute, this was a hacking drill.
00:26:36.880 --> 00:26:39.920
Uh, right, the Claude Mythos, uh, same thing.
00:26:40.000 --> 00:26:47.440
It was the irregular drills with Claude Mythos and the Opus 4.7, um, the anthropic uh you know, sandwich in the park.
00:26:47.680 --> 00:26:49.279
It was told to do those things, right?
00:26:49.359 --> 00:26:55.920
So uh the Hugging Face Open AI one was one of the first major ones, at least that we know about that a company is disclosed.
00:26:56.000 --> 00:26:59.440
Maybe there's tons that just you know companies haven't disclosed.
00:26:59.519 --> 00:27:09.680
But this is the first one where it's like, oh, wait, this AI agent broke its containment and hacked onto you know a website where it wasn't supposed to.
00:27:10.480 --> 00:27:13.599
But that's not the biggest story, right?
00:27:13.759 --> 00:27:18.559
The biggest story is how these agents actually did it, right?
00:27:18.799 --> 00:27:20.720
Which is actually very impressive.
00:27:20.880 --> 00:27:23.839
I mean, it's scary, but it's very impressive, right?
00:27:24.079 --> 00:27:32.079
Uh, so again, this wasn't just one AI, it was a lot of different uh AI agents that were working together.
00:27:32.240 --> 00:27:34.720
Yeah, the AI agents teamed up.
00:27:34.880 --> 00:27:42.799
So, as an example, and this again, you have to tip your hat uh to open AI because they've been talking about this very openly, right?
00:27:42.880 --> 00:27:45.759
They put out a post-mortem report on their website.
00:27:45.839 --> 00:27:56.799
Uh, you know, I think they had two researchers um, you know, at the uh at a recent conference uh talk about this, uh, right, answer questions about this.
00:27:56.960 --> 00:28:11.519
Uh plenty of you know, open AI researchers have been talking about what they learned online, uh, right, where some of the other companies haven't said really anything, much of anything, aside from the original uh admission that it happened.
00:28:11.680 --> 00:28:23.680
Um, but the crazy thing is, one of the agents found that it could leave files for the other inside of Artifactory, which is a developer storage tool that OpenAI reportedly used.
00:28:23.920 --> 00:28:27.039
So that, and this is all the according to reports, right?
00:28:27.119 --> 00:28:28.960
And and and transcripts, all right.
00:28:29.119 --> 00:28:33.359
So that grew into what turned into a hidden message board.
00:28:33.440 --> 00:28:39.200
Uh, and according to a wired report, these AI agents created this hidden message board.
00:28:39.440 --> 00:28:47.519
So the human researchers couldn't find it, and they exchanged reportedly hundreds of thousands of messages to each other.
00:28:47.759 --> 00:28:56.960
So essentially the agents split up the work, they reused the hacks that were working, and they even discussed signing messages to catch an imposter, right?
00:28:57.039 --> 00:28:59.839
They thought that there was an imposter among them, right?
00:29:00.000 --> 00:29:01.039
That's crazy.
00:29:01.200 --> 00:29:08.559
Uh, even more baffling is OpenAI said that they deleted the board, they found it uh July 4th.
00:29:08.640 --> 00:29:13.440
Um, but then the agents rebuilt it in four days and hit Hugging Face again.
00:29:13.680 --> 00:29:16.160
So it right, isn't that crazy?
00:29:16.400 --> 00:29:27.440
After uh, you know, this was kind of erased uh from their memory, they still figured it out by working together and found a way uh to communicate with each other, trying to go behind the human's facts.
00:29:27.599 --> 00:29:30.400
So is that scary?
00:29:31.680 --> 00:29:43.119
Absolutely, um but it is a signal of the capabilities that are ahead.
00:29:43.519 --> 00:29:50.799
And here's why I had a show uh I think it was last week or the week before, right?
00:29:50.960 --> 00:29:59.920
Talking about how we're on the cusp uh probably of entering the RSI phase or the uh recursive self-improvement, right?
00:30:00.000 --> 00:30:15.599
And this is where the models um themselves start making either versions that are already out better, so improving the models themselves, or helping to create or entirely creating new versions of themselves, right?
00:30:15.680 --> 00:30:26.720
So when we think about these agents that have this hacking capabilities and they're already showing tendencies to try to communicate with each other behind humans' backs, right?
00:30:26.880 --> 00:30:34.799
You have to think of what happens when you combine uh that behavior with their ability to improve upon themselves, right?
00:30:34.880 --> 00:30:42.240
Because right now, these agents can run for hours on their own, sometimes even days, and they'll keep retrying, right?
00:30:42.400 --> 00:30:45.200
If we're talking about AI agents that are intentionally, right?
00:30:45.440 --> 00:30:54.799
So now we're pivoting away from the open AI hugging face situation, and we're talking about what happens when these AI agents in the future are intentionally used to do bad, right?
00:30:54.880 --> 00:30:58.240
They're gonna be able to work for hours and communicating with each other, right?
00:30:58.400 --> 00:31:08.319
So one anthropic test that they shared about said that their uh test model scanned about 9,000 real targets after its fake target failed.
00:31:08.640 --> 00:31:22.000
So by itself, any one of these actions might look small, but when you start chaining them together, this leads to a real cyber uh security issue.
00:31:23.039 --> 00:31:25.519
And the real storm, right?
00:31:25.599 --> 00:31:35.920
I I started the show uh talking about that, hey, this you know, AI agent apocalypse, the the AI agent crash, it hasn't happened yet, not even close.
00:31:36.079 --> 00:31:41.759
And I think actually what happens uh or when this will happen is when the open models catch up.
00:31:41.920 --> 00:31:44.000
All right, and here's the reason why, right?
00:31:44.240 --> 00:31:56.319
OpenAI, as an example, they said the hugging face model, that was uh uh a model that was not ready for uh not supposed to go to production, they kind of said they they retired that one, uh, right.
00:31:56.480 --> 00:32:08.079
But for models, proprietary models, like right, with Anthropic, Anthropic had a little scuffle with the US government around you know, some sort of cyber capabilities, and they pulled the model, right?
00:32:08.160 --> 00:32:10.160
And no one in the world could use it.
00:32:10.400 --> 00:32:18.400
But when open weight or open source models have this level of capability and they are very close, right?
00:32:18.480 --> 00:32:21.680
I think Kimmy K3 is the first one that's close.
00:32:21.920 --> 00:32:29.759
I still think the uh, you know, the Astras and the fables of the world are always going to be, you know, one to three months ahead.
00:32:30.000 --> 00:32:41.920
But we are at the point where I'm guessing probably later this year, uh, where now these models are once they're released, they're out.
00:32:42.559 --> 00:32:43.039
Right?
00:32:43.200 --> 00:32:43.920
They're out.
00:32:44.079 --> 00:32:47.279
You can't uh pull an open source model.
00:32:47.599 --> 00:32:49.440
Uh so the difference is, right?
00:32:49.599 --> 00:32:57.440
So the Frontier Labs here in the US, as they release these models, they have them with heavy guardrails.
00:32:57.599 --> 00:33:05.279
Um, and they uh, you know, in theory, have the ability to pull them either via from subscription plans or from APIs.
00:33:05.519 --> 00:33:07.599
Open models are not like that, right?
00:33:07.759 --> 00:33:16.000
You can intentionally uh build on these open models or fork them or deconstruct them to make them less secure, right?
00:33:16.160 --> 00:33:25.599
So when you have these free downloadable models that are only maybe months behind the lockdown ones, that's what I think we as business leaders have to start looking toward.
00:33:25.759 --> 00:33:30.799
Not just what's happening now, and you can look at this and write it off and say, oh, well, you know, they'll shut the model off.
00:33:30.880 --> 00:33:35.680
And, you know, they're working with the, you know, the Trump White House now to make sure these models are safer.
00:33:35.839 --> 00:33:37.680
Yes, they are, um, right.
00:33:37.920 --> 00:33:44.720
But there's always gonna be a case, I think, where the open models, again, a couple months behind.
00:33:44.799 --> 00:33:57.119
And yes, at least right now, you know, the the next, you know, uh open model that comes out that has real bad actor uh AI capabilities, you know, a consumer is not gonna be able to do that.
00:33:57.359 --> 00:34:10.800
But when you talk about bad actors at the state level, uh, right for an adversaries, I would assume uh that these open weight models will be used specifically for those types of bad purposes.
00:34:11.039 --> 00:34:14.159
Um, so there's no amount though, right?
00:34:14.239 --> 00:34:21.440
When we talk about you know senators shaking their fists and you know, people calling for all these slow down, right?
00:34:22.000 --> 00:34:26.000
Once the open models come, it's too late.
00:34:26.239 --> 00:34:27.119
It is too late.
00:34:27.199 --> 00:34:28.800
And yes, we are already getting there.
00:34:28.960 --> 00:34:34.000
We had the little Kimmy uh K3 uh you know instance that we talked about.
00:34:34.079 --> 00:34:49.679
So my assumption is uh open models that are released probably in the fourth quarter might be the ones that we start hearing about in 2027 that start doing some of these things intentionally and are used actually as a weapon to do bad.
00:34:50.239 --> 00:34:58.880
So, my take on this, I think eventually these types of agent hacks or agent crashes will become as common as spam, right?
00:34:59.119 --> 00:35:09.280
So, well, hopefully, what this leads to is better defensive models that become, you know, in the same way that we don't even think about, oh, my email inbox has a spam filter, right?
00:35:09.440 --> 00:35:25.599
I don't know how this is gonna work, but I assume that eventually the the AI labs and the governments and I don't know, uh other uh federal bodies, at least here in the US, are gonna come up with a way to standardize these protections because our businesses will need them, right?
00:35:25.679 --> 00:35:29.360
But I think it is literally gonna become as common as spam, uh, right.
00:35:29.519 --> 00:35:34.239
And and and they'll be like data breaches, robocalls, right?
00:35:34.400 --> 00:35:37.920
These things that we've become accustomed to over the past few decades.
00:35:39.119 --> 00:35:44.800
We're gonna have that level of familiarity, unfortunately, with AI agents going rogue.
00:35:44.960 --> 00:35:47.440
And they're gonna present themselves in many different ways.
00:35:47.519 --> 00:35:53.199
I think the first way that we're gonna see it is ways that many of us don't understand, and that's through cybersecurity exploits, right?
00:35:53.280 --> 00:36:06.559
Exploiting things on systems that we all use, banking systems, softwares that you know, you know, maybe millions of people use through, you know, your company's website, you know, your company's emails, right?
00:36:06.639 --> 00:36:10.079
But again, the consequences will be anything but routine.
00:36:10.320 --> 00:36:24.719
Because when these AI agents can replicate, duplicate, spawn, talk to each other, right, without necessarily humans being able to know what they're up to, uh, this is a lot different than looking at an email and you're like, oh, this looks legitimate.
00:36:24.800 --> 00:36:25.519
Let me click on it.
00:36:25.679 --> 00:36:26.719
Whoops, right?
00:36:26.880 --> 00:36:28.480
Um, much different.
00:36:28.719 --> 00:36:37.760
Because now instead of your account, you know, reforwarding the spam, like if you accidentally clicked on that link and then it sends out the same spam message to everyone in your address book.
00:36:37.920 --> 00:36:42.559
The difference is now your business bank account might be emptied.
00:36:43.679 --> 00:36:47.280
But this is not gonna come with a big splash, right?
00:36:47.440 --> 00:36:53.280
The first AI agent crash that you experience, your company experience, it's gonna be boring, right?
00:36:53.519 --> 00:36:55.920
Uh, it's not gonna be like a blockbuster movie.
00:36:56.000 --> 00:36:57.360
It's just gonna be something boring.
00:36:57.440 --> 00:36:59.280
You may not even notice it at first.
00:36:59.440 --> 00:37:02.960
Uh, it's gonna, you know, leak into your code, your CRM, whatever.
00:37:03.039 --> 00:37:05.119
But it's kind of like that gym agent, right?
00:37:05.280 --> 00:37:11.440
It's gonna, it's it's gonna seem like a routine task that's gonna go past um what was intended.
00:37:11.599 --> 00:37:22.079
So whether someone is attacking you with a rogue agent or the flip side is well, there's something that's more controllable because I think there's a certain element of that, right?
00:37:22.239 --> 00:37:30.400
If it if you are hit with a you know, agent, a rogue AI agent attack, you might not be able to do too much about it now.
00:37:30.639 --> 00:37:35.840
But what you can do now is understanding that this flips both ways, uh, right.
00:37:36.159 --> 00:37:43.360
Because as we're using AI agents in our company, I think sometimes we're doing it haphazardly, right?
00:37:43.440 --> 00:37:53.360
We're just giving it full permissions and having it, you know, read, write, send uh without really any regard given to it.
00:37:53.599 --> 00:38:01.599
So I want us to start thinking about how we can start using AI agents more responsibly and actually pay attention and understand those guardrails.
00:38:01.679 --> 00:38:06.159
Because, like I said, uh these crashing agents, they cut both ways.
00:38:06.400 --> 00:38:09.280
The AI tax, so we have to be ready for that.
00:38:09.440 --> 00:38:12.000
Um, right, but we also have to defend it.
00:38:12.159 --> 00:38:22.559
And the way that we can be better, I think, defensive AI agent users is to make sure that the agents we are deploying do not actually crash the path that we are trying to travel.
00:38:22.719 --> 00:38:30.239
Um, because I think that is actually where most companies are gonna see uh, you know, agent crash happen first, right?
00:38:30.480 --> 00:38:46.320
As these agents can all of a sudden work for hours or days at a time, and we are more and more likely to give them more and more permissions because we're like, wait, this thing can go and update my CRM and this thing can go now.
00:38:46.400 --> 00:38:55.760
And if I just click this YOLO mode button, it can go, you know, take care of all my email responses and it's just gonna, you know, it's it's gonna do so in my voice, right?
00:38:55.920 --> 00:39:02.639
So it's almost like you, as you get more positive capabilities, you let your guard down and then you start to give more and more permissions.
00:39:02.800 --> 00:39:06.239
And I think the human nature is to be a little more lax.
00:39:06.320 --> 00:39:14.239
And part of it, right, and I've seen this in myself, as I get the ability to complete more work, guess what I do?
00:39:14.559 --> 00:39:22.000
I complete more work after I complete more work, uh, right, which I think it is human nature, right?
00:39:22.239 --> 00:39:28.480
As as you start to get you know, the green light on agents, you check a few and you're like, oh yeah, this is good.
00:39:28.639 --> 00:39:31.519
And then you just start sending it in other directions.
00:39:31.599 --> 00:39:41.119
So I think that defenders, we also need to start thinking about how we can use these AI systems to scan logs, catch intrusions, and patch holes fast.
00:39:41.679 --> 00:39:49.280
Because I don't think the future is us versus AI, it's AI attackers versus AI defenders.
00:39:49.679 --> 00:39:50.079
All right.
00:39:50.239 --> 00:39:53.280
So here's your Monday morning playbook for deploying agents safely.
00:39:53.440 --> 00:39:54.960
You need to block the internet by default.
00:39:55.039 --> 00:40:04.800
You need to sit in the same way that researchers do this, you need to set up sandbox and the sandboxes in the right way, you know, when you're testing, especially long-running agents uh with important tasks.
00:40:04.880 --> 00:40:08.159
You need to be able to separate the reading from the doing, uh, right.
00:40:08.239 --> 00:40:16.719
In the same way, if you think of like giving someone access to, you know, your uh a Google document as an example, are they getting read only?
00:40:17.119 --> 00:40:18.639
Can they write and edit?
00:40:18.800 --> 00:40:20.000
Can they leave comments?
00:40:20.159 --> 00:40:21.199
Are they an owner?
00:40:21.360 --> 00:40:21.679
Right.
00:40:21.760 --> 00:40:26.800
You have to think of the same thing when thought uh thinking about your agents and their capabilities, uh, right.
00:40:26.880 --> 00:40:35.679
And at what point um the human, right, the expert-driven loop, the human uh is going in there and making those approvals.
00:40:35.760 --> 00:40:39.199
But you need to give each agent, I think, a short-lived login.
00:40:39.440 --> 00:40:50.320
If you are getting to that access, you start one rung at a time in the same way, think, oh, first they're gonna be view only, then they can be view and comment, then they can be right, then they can be owners.
00:40:50.480 --> 00:40:56.239
You have to think of it in the same way that you might think of sharing a document with an intern, uh, right.
00:40:56.559 --> 00:41:05.119
Um, and you also have to uh make sure that you prioritize traceability and also having a working remote kill switch.
00:41:05.280 --> 00:41:12.159
So yeah, if you get word or wind uh that an AI agent uh is is going rogue, right?
00:41:12.320 --> 00:41:16.719
One you set out to go do good things and now all of a sudden it's doing bad things.
00:41:16.960 --> 00:41:19.840
You can't be like, oh my gosh, it's it's Saturday.
00:41:20.079 --> 00:41:22.639
Uh, you know, I live 20 miles from the office.
00:41:22.719 --> 00:41:25.360
I need bill for my right, I need someone in ops to go in there.
00:41:25.440 --> 00:41:30.239
No, you have to be at any time, you have to know who can click that button.
00:41:30.559 --> 00:41:31.039
All right.
00:41:31.280 --> 00:41:37.599
So as we wrap, I want you to treat this like a fire, not a fire drill.
00:41:37.760 --> 00:41:42.320
Because yes, technically, we are not in the fire, right?
00:41:42.559 --> 00:41:45.760
We are not technically in that crash.
00:41:46.079 --> 00:41:47.360
We are in the warning lap.
00:41:47.440 --> 00:41:51.199
So right now, while we are in the warning lap, you need to treat this as the real thing.
00:41:51.280 --> 00:41:54.480
Because if you don't, by the time the real thing comes, it is going to be too late.
00:41:54.639 --> 00:41:59.199
You need to right now record every agent's full run, every agent that your company has.
00:41:59.360 --> 00:42:03.599
If you don't already have an observability platform, traceability, right?
00:42:03.679 --> 00:42:15.519
If you're using, you know, Microsoft Windows Copilot as an example, intra ID, you need to be able to see and understand at a glance every single action that your agents are taking, uh, right?
00:42:15.679 --> 00:42:18.079
Not just by clicking the chain of thoughts, right?
00:42:18.159 --> 00:42:22.320
You need to be able to observe them and trace all of their steps.
00:42:22.960 --> 00:42:26.559
You should never give an agent more power than you can watch.
00:42:26.719 --> 00:42:30.400
You need to be able to undo and survive any action that it takes.
00:42:30.559 --> 00:42:46.000
Uh, and lastly, I think the companies that are preparing for this now and putting these steps into place, they're gonna be the ones that are most protected when the AI agent crash actually starts happening because it has not started yet, y'all, but it is coming soon.
00:42:46.079 --> 00:42:52.559
And that's not me being like uh, you know, a doomsdayer or a crazy, like, oh, watch out, right?
00:42:52.719 --> 00:42:54.880
If you listen to this show, I'm not like that.
00:42:54.960 --> 00:42:55.519
Uh, right.
00:42:55.599 --> 00:42:59.679
I I try to uh ground myself in practically what's happening.
00:42:59.840 --> 00:43:04.639
And what's happening now is these agents are becoming more and more capability.
00:43:04.800 --> 00:43:08.079
Uh sorry, these agents are becoming more and more capable.
00:43:08.159 --> 00:43:13.519
Um, and the capabilities themselves are compounding quickly, right?
00:43:13.679 --> 00:43:24.800
And that means, yes, uh, you know, one of the biggest discussions uh, you know, in AI and now in Washington right now is building in these safeguards and protections.
00:43:25.119 --> 00:43:31.599
But what we have to keep an eye on is when the open models match where we're at today.
00:43:32.000 --> 00:43:50.719
Because all the pausing and guardrails and pacing in the world doesn't mean a thing if there is a Chinese open model in four months that has mythos or astra level capabilities, because at that point, the gloves are off and we all have to be ready.
00:43:50.880 --> 00:43:51.199
All right.
00:43:51.280 --> 00:43:58.639
I hope this one was helpful going over rogue AI agents, why breakouts are happening more, and how companies should prepare.
00:43:58.719 --> 00:44:02.239
If this was helpful, do me a favor, subscribe if you're listening on the podcast.
00:44:02.400 --> 00:44:06.719
Then go to your everydayai.com, sign up for the free daily newsletter.
00:44:06.800 --> 00:44:07.360
Thanks for tuning in.
00:44:07.599 --> 00:44:10.000
See you back tomorrow and every day for more everyday AI.
00:44:10.079 --> 00:44:11.119
Thanks, y'all.
