WEBVTT
00:00:01.040 --> 00:00:03.680
There's always been AI risk.
00:00:03.919 --> 00:00:18.960
But in the early days of large language models and chatbots, that risk was like Bill getting something wrong in the blog post or Deborah putting up a hallucinated stat in the onboarding guide that was littered with M dashes and delves.
00:00:19.600 --> 00:00:26.239
But AI Risk Today is a legit different ball game than risk was three and a half years ago.
00:00:26.399 --> 00:00:31.519
I mean, heck, AI Risk today is unrecognizable from what it was three and a half months ago.
00:00:31.679 --> 00:00:33.439
And that's not an exaggeration.
00:00:33.679 --> 00:00:41.119
Because after hearing for like five years that we're six months away from real AI agents, well, it finally happened.
00:00:41.280 --> 00:00:50.799
And it was actually this perfect storm of multiple events that led to an unexpected business scenario that the corporate world now faces today.
00:00:50.960 --> 00:00:58.640
You either get on board with AI agents quickly or get left behind, but do it too quickly, and you could go under.
00:00:59.200 --> 00:01:02.640
The risk, the security, and the sprawl are real.
00:01:02.799 --> 00:01:07.840
And so we're gonna tackle it all today on Everyday AI.
00:01:08.000 --> 00:01:10.000
The start here, series edition.
00:01:10.239 --> 00:01:12.079
All right, well, welcome to Everyday AI.
00:01:12.159 --> 00:01:13.280
My name's Jordan Wilson.
00:01:13.359 --> 00:01:14.959
And if you're new here, this thing's for you.
00:01:15.040 --> 00:01:23.840
It's your daily live stream podcast and free daily newsletter helping everyday business leaders like you and me keep up with all the news, like a thousand new agents a day.
00:01:23.920 --> 00:01:24.400
What do we do?
00:01:24.480 --> 00:01:25.040
What do we try?
00:01:25.120 --> 00:01:30.159
Well, tune in, I tell you, and uh help you make the right decisions to grow your company and your career.
00:01:30.319 --> 00:01:36.159
So after 700 plus episodes, I realized I couldn't answer the most common question that people had for me.
00:01:36.239 --> 00:01:37.920
Like Jordan, you have a lot of episodes.
00:01:38.000 --> 00:01:38.719
Where do I start?
00:01:38.879 --> 00:01:41.920
Well, that's why I created the Start Here series.
00:01:42.000 --> 00:01:49.439
So the Start Here series is the essential podcast series to both learn the AI basics and to double down on your knowledge.
00:01:49.599 --> 00:01:59.439
Uh so this is volume nine, and you can go listen to all of the uh episodes in our Start Here series if you just go to starthearseries.com.
00:01:59.519 --> 00:02:08.159
So that will give you uh free access to our inner circle community, and it'll put you right in the Start Here Series space.
00:02:08.319 --> 00:02:15.039
So you can go uh listen to all of them, watch them, read about them all in one place, interact with others who are doing the same.
00:02:15.199 --> 00:02:19.680
All right, so uh one other thing that you need to do before we get started.
00:02:19.919 --> 00:02:23.919
My gosh, y'all, you you have to go listen to these.
00:02:24.000 --> 00:02:26.960
Uh, episode 712 and 713.
00:02:27.120 --> 00:02:30.159
That is our 2026 AI prediction and roadmap series.
00:02:30.240 --> 00:02:38.800
That's like a culmination of a thousand hours of work over the years uh or over the past year uh to give you guys the blueprint for 2026.
00:02:38.879 --> 00:02:40.879
So make sure you go listen to those.
00:02:41.120 --> 00:02:41.520
All right.
00:02:41.680 --> 00:02:46.080
Uh, if this sounds kind of like our last episode in the Start Here series, not exactly.
00:02:46.159 --> 00:02:48.479
Uh so make sure you go listen to that one if you didn't already.
00:02:48.560 --> 00:02:55.280
So this is more of just the state of AI agents, where they are, what they are, how we should use them, should we use them.
00:02:55.439 --> 00:02:58.639
So make sure you go listen to uh volume eight from yesterday.
00:02:58.800 --> 00:03:00.479
That's episode 717.
00:03:00.879 --> 00:03:07.759
But today we're here to talk about the other side, the risk, the security, and the sprawl that's changing everything.
00:03:08.080 --> 00:03:14.879
So here's what we're gonna be covering on today's show and well, why it urgently matters.
00:03:15.199 --> 00:03:20.319
Because AI models didn't just get smarter, they got hands, right?
00:03:20.400 --> 00:03:29.199
Uh the risk model changed when AI moved from generating text like it was three and a half years ago, to now it's taking real actions.
00:03:29.280 --> 00:03:32.560
And a lot of times actions we're not aware of, and that's the scary part.
00:03:32.719 --> 00:03:43.919
And an agent connected to your email and calendar or you know, your company's data can act fast, confident, and wrong in the same way that AI models hallucinated three years ago.
00:03:44.080 --> 00:03:45.680
Well, they can still hallucinate now.
00:03:45.840 --> 00:03:48.560
And this isn't a future concern, this is happening right now.
00:03:48.719 --> 00:03:55.919
So on today's show, uh, we're gonna go over a simple mental model for why agent risk is fundamentally different from chatbot risk.
00:03:56.000 --> 00:04:02.479
Uh, we're gonna talk about what OpenAI, Google, Anthropic, and Microsoft are actually building to address this risk.
00:04:02.639 --> 00:04:11.599
And I'm gonna give you a practical Monday morning playbook that you can start using this week to address the risk security and sprawl.
00:04:11.919 --> 00:04:12.800
Sound good?
00:04:13.120 --> 00:04:15.120
Yeah, sounds good to me.
00:04:15.199 --> 00:04:17.839
I'm I'm excited and I wrote all this.
00:04:18.000 --> 00:04:21.279
All right, so let's get a quick little catch up here.
00:04:21.439 --> 00:04:26.480
Um, so probably if you're listening to this show, AI is not new to you, right?
00:04:26.639 --> 00:04:30.560
But um, let me just give everyone the the the briefer here, right?
00:04:30.720 --> 00:04:39.279
So essentially from 2022 to mid-2023, early 2024, large language models were largely text systems, right?
00:04:39.439 --> 00:04:47.040
With the the risk was just limited to misinformation and data leaks and you know, using hallucinations and looking foolish.
00:04:47.199 --> 00:04:51.279
Uh, but that started to change, I'd say, in the mid-2025.
00:04:51.439 --> 00:04:58.160
Uh, so that's when the models started getting well exponentially more capable and agentic by default.
00:04:58.319 --> 00:05:01.120
That's the thing that people don't understand.
00:05:01.279 --> 00:05:01.680
Uh, right.
00:05:01.759 --> 00:05:11.759
I've I've actually had two great conversations with uh two really smart minds, uh, kind of the head of um agents at Cloudflare and then uh the head of Microsoft Research.
00:05:11.839 --> 00:05:15.600
And I learned a lot by talking to them both, you know, before and after the show as well.
00:05:15.759 --> 00:05:21.839
Um but one thing that kind of came true is well, no one really knows what constitutes an agent and what doesn't.
00:05:22.000 --> 00:05:28.720
But I think that we can all agree that even today's large language models, they're agentic by nature, right?
00:05:28.879 --> 00:05:33.439
They, well, a lot of us are giving them access to all of our data.
00:05:33.600 --> 00:05:39.279
Um, and not in all cases do they have right access, but in many cases, they do.
00:05:39.519 --> 00:05:47.600
Um, so when these models can think and act on their own and um, you know, spin up a virtual environment and a terminal and access your computer, right?
00:05:47.759 --> 00:05:48.639
I keep saying this.
00:05:48.800 --> 00:05:53.519
Even right now, I have multiple agents running on my computer, right?
00:05:53.680 --> 00:05:56.480
I have codecs and clawed code going right now.
00:05:56.560 --> 00:05:59.920
I'll probably spin up uh anti-gravity uh later.
00:06:00.319 --> 00:06:06.240
I always have agents running and they have access, and I don't necessarily know what they're doing uh, you know, in between.
00:06:06.319 --> 00:06:07.839
I always go back and look when they're done.
00:06:08.079 --> 00:06:14.319
But that's what's really changed um in mid uh, you know, 2025.
00:06:14.480 --> 00:06:34.319
But here we are in early 2026, and this is where uh now we're gonna start talking about and actually putting into practice all those buzzwords, uh, right, that people just you know started chatting about in like 2024 to you know sound super smart, like oh, governance and audit logs and you know, uh isolation uh protocols, right?
00:06:34.399 --> 00:06:35.680
Like all those well, okay.
00:06:35.759 --> 00:06:40.959
Well, good thing that everyone was talking about it, uh, because you know, we were trying to get our our buzzword bingo.
00:06:41.040 --> 00:06:42.879
Um well, now we actually need it, right?
00:06:42.959 --> 00:06:50.160
Now it's uh time to talk about uh ethics and govern governance and guardrails when it comes to uh agentic capabilities.
00:06:55.279 --> 00:06:58.480
I like to put it like this to keep it very simple.
00:06:59.759 --> 00:07:09.839
Um 2022 and before, I'd say that AI was a dumb stationary brain, but it was a brain, right?
00:07:09.920 --> 00:07:12.079
So everyone was blown away, like, oh my gosh, it can think.
00:07:12.240 --> 00:07:13.120
Well, it was dumb.
00:07:13.279 --> 00:07:17.279
It was stationary, didn't move, couldn't really think ahead, right?
00:07:17.439 --> 00:07:22.480
And in 2023, well, it became a dumb stationary brain with tools, right?
00:07:22.560 --> 00:07:29.600
That's when uh, you know, the early version of uh ChatGPT plus uh, you know, GPT 4, it had tools, right?
00:07:29.680 --> 00:07:31.439
It could go on the internet as an example.
00:07:31.519 --> 00:07:37.360
So that's when it uh really started to open up what it could do, or at least the information that it could access.
00:07:37.600 --> 00:07:44.480
In 2024, uh, well, it was still a stationary brain with tools, but it went from a dumb brain to a smart brain.
00:07:44.560 --> 00:07:51.759
I would say in 2024 was the first time that we actually had smart models, uh, because at the end of 2024, that's when we got reasoning models.
00:07:51.839 --> 00:08:02.399
Uh and then in 2025, I think we still had smart brains with tools, but the difference now, instead of it being a stationary brain, it was a proactive brain in 2025.
00:08:02.639 --> 00:08:10.879
It could go out, at least, especially at the end of 2025, the second and third quarter, it can make moves on its own, right?
00:08:11.439 --> 00:08:12.240
Proactively.
00:08:12.800 --> 00:08:18.720
It could schedule things and or just you know, they can start acting over long periods.
00:08:18.959 --> 00:08:28.319
And then what brings it to 2026 is well, now we have that smart, proactive brain with tools and arms, right?
00:08:28.720 --> 00:08:36.240
So the tools are cool, but when you have arms, you can actually use them in a real way.
00:08:36.320 --> 00:08:44.559
And I think that's where you know agents now have teeth when they are autonomous, uh proactive, and smart.
00:08:45.279 --> 00:08:47.919
So here's why agent risk I think feels different.
00:08:48.159 --> 00:08:53.600
Um, because you know, like I said, with the chatbots, there was always risk, but was it really?
00:08:54.000 --> 00:09:00.159
I mean, yeah, worst thing you can do is you get in trouble putting out something hallucinated and you look foolish, right?
00:09:00.399 --> 00:09:02.000
Uh does your company go under?
00:09:02.159 --> 00:09:03.039
Probably not.
00:09:03.279 --> 00:09:03.840
Uh right.
00:09:03.919 --> 00:09:06.480
Do you expose every single dark sea, right?
00:09:06.639 --> 00:09:07.600
It's not that bad.
00:09:07.840 --> 00:09:09.440
I mean, it is, but it's not.
00:09:09.600 --> 00:09:16.639
But that the the new agent layer is just a whole new type of risk that we're not even ready for.
00:09:16.799 --> 00:09:20.240
If I'm being honest, if you listen to the show, I say this a lot.
00:09:20.399 --> 00:09:22.720
I'm like, this year is going to be scary.
00:09:23.279 --> 00:09:27.200
People are not ready, businesses are not ready, and I mean that, right?
00:09:27.440 --> 00:09:32.399
That kind of business predicament that I talked about in the opening of the show there, that is real.
00:09:33.120 --> 00:09:39.679
If you don't run to use AI agents this year, you're toast.
00:09:40.399 --> 00:09:41.279
You are literally toast.
00:09:41.840 --> 00:09:50.159
I don't care if you're a uh small business or a$20 billion revenue business, you're toast.
00:09:50.480 --> 00:09:50.799
All right.
00:09:51.679 --> 00:09:54.879
But if you sprint too quickly, you can go under, right?
00:09:55.279 --> 00:10:02.960
Because you could make devastating mistakes that that would be uh highly improbable to recover from, right?
00:10:03.120 --> 00:10:09.440
Because AI agents can do things uh much worse than even the worst human, right?
00:10:09.600 --> 00:10:12.320
Have you ever heard stories or maybe you've experienced this, right?
00:10:12.480 --> 00:10:16.000
Kind of a rogue employee happens too often, right?
00:10:16.080 --> 00:10:21.840
Someone's someone's bitter maybe about getting fired or not getting a promotion, and they do something absolutely crazy, right?
00:10:21.919 --> 00:10:26.159
Maybe expose all the company secrets or you know, release some files.
00:10:26.320 --> 00:10:26.879
I don't know.
00:10:27.600 --> 00:10:29.440
Okay, that's a human, right?
00:10:29.600 --> 00:10:32.559
That's a single human, and you can see that human.
00:10:32.639 --> 00:10:34.639
You have eyes on that human, right?
00:10:34.879 --> 00:10:37.200
Bill and IT is watching that human.
00:10:38.080 --> 00:10:39.039
Agents are different.
00:10:40.000 --> 00:10:46.240
Agents move a hundred, a thousand times faster than that one person, but you can't always see agents.
00:10:46.639 --> 00:10:47.200
And guess what?
00:10:47.279 --> 00:10:49.440
That one disgruntled employee, that's one.
00:10:50.240 --> 00:10:51.200
This isn't a video game.
00:10:51.279 --> 00:10:52.480
You can't respond 10 times.
00:10:52.639 --> 00:10:53.360
Agents can.
00:10:54.000 --> 00:11:00.240
Agents can spawn sub-agents like that, and those sub-agents can spawn like that, right?
00:11:00.399 --> 00:11:08.720
So think of in the way like a virus might spread across your computer or across the human body, uh, right, and replicate and duplicate.
00:11:08.879 --> 00:11:10.480
It's the same thing with AI agents.
00:11:10.639 --> 00:11:12.080
A rogue employee can't do that.
00:11:12.159 --> 00:11:14.639
And that's why this risk is very different.
00:11:14.720 --> 00:11:16.000
It is very real.
00:11:16.559 --> 00:11:25.919
So I think that we spend too much time uh thinking about the positives and the optimistic side of AI agents, which is great, right?
00:11:26.080 --> 00:11:33.440
Yes, oh now all of a sudden I have, you know, 320 uh AI agents, you know, doing all my work for me around the clock.
00:11:33.519 --> 00:11:34.320
Oh, that's cool.
00:11:34.559 --> 00:11:34.960
All right.
00:11:35.200 --> 00:11:38.240
But what about the risk, the security, and the sprawl?
00:11:39.440 --> 00:11:46.080
And more teams are experimenting now than ever before because no one's got it all figured out.
00:11:46.240 --> 00:11:52.240
And that means more sprawl uh and more exposure, uh, especially if you don't have guardrails up.
00:11:52.399 --> 00:11:56.080
So here's essentially the three surfaces where agent risk actually lives.
00:11:56.240 --> 00:11:57.360
Number one is the input.
00:11:57.519 --> 00:12:03.840
That's kind of that untrusted content that can contain hidden instructions, agents you know, treat like real commands.
00:12:04.000 --> 00:12:07.759
All right, that I don't think is going to be as big of a deal, right?
00:12:07.919 --> 00:12:11.919
Your inputs, don't get me wrong, things can go wrong, right?
00:12:12.080 --> 00:12:14.240
People are blindly copying and pasting.
00:12:14.320 --> 00:12:19.679
So you can copy and paste prompt injections, but prompt injections are the big thing, and we're gonna talk about that here in a second.
00:12:19.759 --> 00:12:22.559
But still, inputs, they can be poisoned, right?
00:12:22.639 --> 00:12:24.559
There's there's things that you might not know.
00:12:24.720 --> 00:12:33.679
And in the same way, a rogue human can create a lot of rogue AI agents that can create a lot of risk and a lot of sprawl that's gonna be uncontrollable.
00:12:33.759 --> 00:12:42.000
So, you know, inputs, I think you really only have to worry about it, you know, if someone really doesn't know what they're doing or if someone is trying to be malicious, which again, those things are gonna come up.
00:12:42.159 --> 00:12:46.960
But think of that one bad employee, what they can now do if they know agentic AI, right?
00:12:47.120 --> 00:12:47.519
Yeah.
00:12:47.679 --> 00:12:58.480
Uh talk about malware, spy spyware, uh, right, ransomware that we, you know, these companies that have paid, you know, millions of dollars, hundreds, tens of millions of dollars for ransomware.
00:12:58.559 --> 00:13:00.240
It's gonna be way worse with agents.
00:13:00.320 --> 00:13:02.159
Uh, so the second layer is tools.
00:13:02.240 --> 00:13:09.200
And I think this is where it starts to get a little dicey in terms of, well, the capabilities are wild, right?
00:13:09.360 --> 00:13:15.600
So every permission and connector that you add expands that blast radius essentially when something goes wrong.
00:13:15.759 --> 00:13:19.519
In the same way that I walked you guys through the, oh, it's a dumb stationary brain.
00:13:19.679 --> 00:13:23.440
But hey, once that brain gets tools, okay, now it can start doing a lot of things.
00:13:23.600 --> 00:13:26.799
And the same thing on the agent gone wrong side, right?
00:13:26.960 --> 00:13:33.679
If it was a dumb stationary uh agent with no tools and no arms, it's like, all right, well, have fun, buddy.
00:13:33.759 --> 00:13:34.000
Right?
00:13:34.080 --> 00:13:36.559
You're in a glass case of emotion just shaking up.
00:13:36.720 --> 00:13:39.360
Uh when they have tools, that's where things go wrong, right?
00:13:39.519 --> 00:13:43.679
When they have uh access to your terminal, right?
00:13:43.759 --> 00:13:45.919
To your computer terminal, that's where things can go wrong.
00:13:46.080 --> 00:13:49.279
When they can run code on a machine, that's where things can go very wrong.
00:13:49.440 --> 00:13:53.519
And then last but not least, and this is the big one, this is actions.
00:13:53.759 --> 00:14:04.720
And this is the biggest thing that um if I had to boil down the biggest change in risk and why it matters now more than ever, it's outputs to actions, right?
00:14:04.879 --> 00:14:09.679
What we had to worry about a couple of years ago from AI in terms of risk was the outputs.
00:14:09.840 --> 00:14:16.799
Now we have to worry about the actions, but it's actions at scale and actions that we might not even necessarily be able to see, right?
00:14:17.039 --> 00:14:21.840
Silent, unintended workflows that you may not be tracing tracing.
00:14:22.879 --> 00:14:33.360
And really what this comes down to, well, it's this combination of increased capabilities from agents and moving in the shadows, and that's an enterprise nightmare.
00:14:33.440 --> 00:14:36.080
That is literally the formula for an enterprise nightmare.
00:14:36.240 --> 00:14:38.000
So, some stats here for you, right?
00:14:38.080 --> 00:14:44.159
So, right now, 57% of employees uh at least admit to using personal AI accounts for work.
00:14:44.240 --> 00:14:45.759
It's way more than that, let's be honest.
00:14:45.840 --> 00:14:47.200
That's just the number that admit.
00:14:47.279 --> 00:14:51.919
Uh, a third admit to inputting sensitive data into unapproved tools, right?
00:14:52.000 --> 00:14:54.159
So your shadow AI use case there.
00:14:54.320 --> 00:14:59.840
And here's the thing you can't govern what you can't see.
00:15:00.720 --> 00:15:06.480
And most organizations can't see their agent footprint at scale.
00:15:08.639 --> 00:15:12.320
This is what I call the three types of dark AI.
00:15:12.480 --> 00:15:12.799
All right.
00:15:12.879 --> 00:15:22.480
For the most part, you're not gonna find you know the three dark types of or the three types of dark AI online is something that I've kind of uh categorized them in.
00:15:22.559 --> 00:15:28.080
But I think it's really helpful uh to get a better glimpse and a better understanding of the categories of risk.
00:15:28.240 --> 00:15:31.279
So, number one, we all know this, shadow AI, right?
00:15:31.360 --> 00:15:35.679
That's just essentially unapproved or unknown AI use, right?
00:15:35.919 --> 00:15:45.600
If you've been using, you know, Chat GPT on your personal computer uh because Copilot is the AI that's approved, but you want to use Chad GPT and you copy and paste things over as an example, right?
00:15:45.679 --> 00:15:47.679
That's shadow AI, but that's been around.
00:15:47.759 --> 00:15:49.039
Everyone knows that, right?
00:15:49.200 --> 00:15:54.480
Uh what you've maybe heard of, maybe not, it's the next kind of tier, and that's agent sprawl.
00:15:54.720 --> 00:15:58.080
But agent sprawl is known, right?
00:15:58.240 --> 00:16:06.159
So that's essentially when you have approved agents, but you're not sure how to wrangle them or observe them all.
00:16:06.240 --> 00:16:15.440
You're like, oh, well, yeah, Bill gave us that, you know, agent to help with finances, but we're not really sure what it's doing, right?
00:16:15.919 --> 00:16:18.000
Um, we think it's doing good, right?
00:16:18.080 --> 00:16:20.879
I checked the outputs, but I don't really know how it's getting there.
00:16:21.120 --> 00:16:23.840
That's that's that's the beginning of agent sprawl.
00:16:24.000 --> 00:16:27.759
But the thing is, agent sprawl goes quickly.
00:16:28.000 --> 00:16:34.639
In the same way, a snowball at the top of the mountain might come rumbling down a thousand times the size.
00:16:34.799 --> 00:16:38.879
That is where we get into then dark agent sprawl.
00:16:39.120 --> 00:16:39.759
All right.
00:16:40.080 --> 00:16:44.879
This sounds like a like a screen name for uh for aim back in the back in the 90s.
00:16:45.120 --> 00:16:46.080
You guys remember that?
00:16:46.240 --> 00:16:46.559
Right?
00:16:46.720 --> 00:16:49.679
AIM, uh, you know, instant messenger, right?
00:16:49.919 --> 00:16:53.919
Dark Agent Sprawl, uh 2026.
00:16:54.000 --> 00:16:59.519
Uh that's gonna be my uh my username on the uh everyday AI inner circle community.
00:16:59.679 --> 00:17:05.359
Um but essentially that's where there's well dark agent sprawl can be a couple of things because agent sprawl, it's a problem.
00:17:05.440 --> 00:17:07.839
But like I said, for the most part, that's a known risk.
00:17:08.000 --> 00:17:11.759
And you're like, yeah, we have these agents going all over the place.
00:17:12.240 --> 00:17:18.480
We we have no guardrails, we have no uh traceability, observability, right?
00:17:18.720 --> 00:17:20.559
It's but but you know of it.
00:17:20.640 --> 00:17:30.400
Dark agent sprawl is agents you don't know about, those are unapproved agents that are working in or on your company, and you can't observe them because you don't know about them.
00:17:30.480 --> 00:17:35.680
So this is well, this can be shadow AI, you know, uh gone agency, right?
00:17:35.759 --> 00:17:41.839
So people plugging in agents that aren't approved and company doesn't know, but there's also another kind, right?
00:17:42.160 --> 00:17:46.880
Uh and I think we're gonna see a lot of this, maybe not in 2026, but in 2027.
00:17:46.960 --> 00:17:52.559
That's going to be the equivalent of malware or spyware, but agents, right?
00:17:52.640 --> 00:18:07.440
People sending uh and seeding agents out specifically in the same way that they would for spyware, uh, malware, ransomware to make money, uh, right, to extract uh, you know, value from businesses.
00:18:07.519 --> 00:18:08.960
And that's what's gonna happen.
00:18:09.119 --> 00:18:11.200
Um, that's the next version of this.
00:18:11.279 --> 00:18:14.319
That's not necessarily dark agent sprawl, but there's two sides, right?
00:18:14.480 --> 00:18:17.359
Dark agent sprawl can start out innocent enough, uh, right?
00:18:17.519 --> 00:18:21.519
Oh my gosh, you know, our company's not approving, you know, any of my AI agents.
00:18:21.599 --> 00:18:22.559
Well, okay, whatever.
00:18:22.720 --> 00:18:32.160
I'm gonna go ahead and um, you know, unleash uh, you know, clawed code and get, you know, 50 instances of clawed code going up and they're all gonna spin their agents.
00:18:32.240 --> 00:18:34.960
Well, one person might know, but the rest of the company is in the dark.
00:18:35.119 --> 00:18:38.480
And those agents could replicate, duplicate, and you can't observe them.
00:18:38.559 --> 00:18:43.680
But the other thing is, well, bad actors, dark agent sprawl, that's a thing too.
00:18:44.079 --> 00:18:46.480
So why is this all happening now?
00:18:46.720 --> 00:18:50.240
All these risks and security concerns and the sprawl, why why now?
00:18:50.400 --> 00:18:54.400
Because, like I've said, we've been six months away for five years, right?
00:18:54.480 --> 00:18:59.440
We're six months away from great agents, but then it's all it almost seems like we didn't get a six-month warning.
00:18:59.680 --> 00:19:01.440
They just popped up, right?
00:19:01.759 --> 00:19:03.759
Like December 2025, right?
00:19:03.839 --> 00:19:08.799
Like everyone was winding down for you know the holidays, you know, especially here in the US.
00:19:08.880 --> 00:19:12.720
And then you go back to work in January, you're like, what the frick happened, right?
00:19:13.200 --> 00:19:17.279
You're like, it's like, whoa, whoa, wait, the the agents are actually here now.
00:19:17.359 --> 00:19:18.559
Where's the six-month warning?
00:19:18.720 --> 00:19:21.519
Why don't we get get word of this in in June or July?
00:19:21.680 --> 00:19:23.359
But there's three reasons why.
00:19:23.440 --> 00:19:31.599
And it is literally the perfect storm, not just the ingredients happening, but at the exact right or wrong time.
00:19:31.759 --> 00:19:32.000
All right.
00:19:32.079 --> 00:19:40.000
So, number one, the number one element uh that was mixed in the bowl and it's exploding is the reasoning threshold.
00:19:40.400 --> 00:19:40.880
Right.
00:19:41.039 --> 00:19:58.799
So improved reasonings from models like you know, GPT-5.3.1, uh, their new model, very impressive, and you know, Opus Sonnet 4.6 from Anthropic, these are all built to be agent native.
00:19:59.279 --> 00:20:01.519
That's how they build them now, right?
00:20:01.759 --> 00:20:06.480
They're not building models that are great at you know, reading, writing, and comprehension first.
00:20:06.880 --> 00:20:11.680
No, what they're focusing on is is the harness and and the tool use.
00:20:11.759 --> 00:20:12.799
That's what's first, right?
00:20:12.880 --> 00:20:19.759
Even you know, Google's um, you know, their new model yesterday, everything that they highlighted was tool use, right?
00:20:19.920 --> 00:20:27.359
And and and how, you know, improved tool use and the scaffolding there has, you know, really helped them um improve their outputs.
00:20:27.839 --> 00:20:29.119
But that's that's the thing.
00:20:29.279 --> 00:20:32.480
These models, their reasoning ability is legit through the roof.
00:20:32.559 --> 00:20:39.519
They can plan steps ahead, uh, self-correct errors, and move beyond uh reactive behavior into proactive.
00:20:39.599 --> 00:20:43.839
And that's kind of reduced agent reliability from around 50% to 90%.
00:20:44.240 --> 00:20:46.880
And that's perfect storm right there, right?
00:20:47.039 --> 00:20:49.599
50% coin flip, you're not doing that, right?
00:20:49.839 --> 00:20:52.880
Would you hire an employee that has a 50% chance to fail right away?
00:20:53.039 --> 00:20:54.960
Probably not, but 90%, okay.
00:20:55.119 --> 00:20:56.960
That's an A employee.
00:20:57.519 --> 00:21:00.640
Um, step two, computer use improvements.
00:21:00.720 --> 00:21:02.880
This seems small, but this is big.
00:21:03.039 --> 00:21:03.359
Okay.
00:21:03.759 --> 00:21:06.559
The models have to be insanely smart, right?
00:21:06.640 --> 00:21:08.079
They have to be able to think, reason.
00:21:08.160 --> 00:21:11.440
They have to have genius level IQ, which is what today's models do, right?
00:21:11.519 --> 00:21:17.519
If you look at offline IQ tests, they're scoring in the genius level, smarter than 99.9% of people, right?
00:21:17.599 --> 00:21:19.200
Uh, but the computer use is important.
00:21:19.279 --> 00:21:25.759
It doesn't matter if it can't go and use a computer, but because at least for now, that's how businesses generate value.
00:21:25.839 --> 00:21:29.920
I think eventually value will be just generated agentically, right?
00:21:30.079 --> 00:21:32.319
The web has gone agentic, right?
00:21:32.480 --> 00:21:39.920
Google and Microsoft uh announced support for essentially an MCP version of the web, um, where websites can talk to each other, agents can talk to each other.
00:21:40.079 --> 00:21:43.279
So we'll see how you know business value is extracted in the future.
00:21:43.519 --> 00:21:49.119
Because a lot of times it's been through the website stack, the software stack, but you know, who knows what it'll be in the future?
00:21:49.200 --> 00:21:53.200
But right now, agents can use computers better than humans, right?
00:21:53.359 --> 00:21:59.839
So uh computer use means uh these models can use a mouse computer, they can uh click, they can use API.
00:22:00.480 --> 00:22:01.680
They can talk to each other.
00:22:02.000 --> 00:22:05.920
But the big thing is the computer use capability gap has essentially been solved.
00:22:06.240 --> 00:22:10.319
So Claude Sonnet 4.6, which came out well earlier this week.
00:22:10.480 --> 00:22:10.799
All right.
00:22:10.960 --> 00:22:14.960
So if you're listening to this in six months, right, it came out in uh mid-February.
00:22:15.039 --> 00:22:19.920
So imagine in six months, uh, these these uh computer use models are gonna be uh insanely good.
00:22:20.000 --> 00:22:26.079
Uh but it's the first um the first one that scored better than humans, right?
00:22:26.160 --> 00:22:37.759
So 72.5 uh success rate on the OS World benchmark, and that's surpassing human performance for the first time and nearly quintupling the scores of 2024.
00:22:37.839 --> 00:22:38.240
That's the thing.
00:22:38.319 --> 00:22:48.559
And I think that's one of the reasons why, you know, even though the uh maybe the foundation was there for AI agents uh, you know, a year, year and a half ago, they couldn't use the computer.
00:22:48.640 --> 00:22:50.480
It was so slow, uh, right?
00:22:50.640 --> 00:22:53.759
It was, you know, very, very, very slow computer vision, right?
00:22:53.839 --> 00:22:57.279
Every time, if you wanted to click something, it would take three minutes.
00:22:57.440 --> 00:22:57.759
Okay.
00:22:58.079 --> 00:23:08.960
Now models can go as fast as humans, or at least, you know, uh not just clicking and navigating interfaces, but just uh a range of computer use tasks.
00:23:09.599 --> 00:23:15.839
And then the third thing that has created this perfect agentic storm, well, it's the context window in memory, uh, right.
00:23:15.920 --> 00:23:20.319
So being able to work on tasks for a long period of time, uh right.
00:23:20.480 --> 00:23:30.319
I think my uh my record uh so far, I think I got to about 10 hours uh overnight the other night on codex, um, which was really fun to do, right?
00:23:30.559 --> 00:23:34.160
But it kept its memory persistent the whole time, right?
00:23:34.319 --> 00:23:36.559
It didn't forget what I told it when I went to wake up.
00:23:36.720 --> 00:23:39.839
You know, when I woke up, yeah, I spent time reading through the chain of thought.
00:23:39.920 --> 00:23:40.799
And I'm like, oh, sweet.
00:23:40.880 --> 00:23:48.720
Not only did it remember anything, but I'm going through, I'm, you know, uh tracing it, I'm, you know, making sure it kind of stayed within the confines of the instructions I gave it.
00:23:48.799 --> 00:23:49.279
And it did.
00:23:49.359 --> 00:23:51.039
But, you know, that's another reason why it's happening.
00:23:51.119 --> 00:24:13.519
And y'all, let me tell you, if you haven't in the last like three, four weeks, if you haven't gone out and used ChatGPT or you know, OpenAI's codex, if you haven't used um, you know, Claude Code or Claude Cowork from Anthropic, if you haven't used uh Anti-Gravity uh from Google, I'm not saying this to like be the guy, you you're gonna get left behind, right?
00:24:13.680 --> 00:24:16.240
And if you're listening to this show, I don't want you left behind.
00:24:16.319 --> 00:24:19.039
I'm actually thinking, all right, I don't know this yet.
00:24:19.200 --> 00:24:24.640
It might might kind of um get a little bit of a vibe working focus.
00:24:24.880 --> 00:24:27.759
Um, you you know, here in the uh first or second quarter.
00:24:27.839 --> 00:24:32.319
So uh make sure if you're not already in our uh inner circle community, get in there.
00:24:32.480 --> 00:24:35.839
Um yeah, I'm gonna start putting putting some stuff in there.
00:24:35.920 --> 00:24:38.160
Just keep keep uh keep your eyes open for that.
00:24:38.480 --> 00:24:43.279
All right, but let's keep going a little bit here as we uh get gonna be wrapping up here in a minute.
00:24:43.359 --> 00:24:52.400
Uh, let's talk a little bit about how the biggest AI companies are actually responding right now because the risk is well known because it has literally been an explosion and it is going to get worse.
00:24:52.559 --> 00:24:57.200
So I want to talk about a little bit about kind of the uh big picture focus of the big four labs.
00:24:57.279 --> 00:25:01.279
So right now, OpenAI is taking more of a human approval uh approach, right?
00:25:01.440 --> 00:25:05.119
So Codecs is kind of their command center uh to review agent decisions.
00:25:05.359 --> 00:25:22.400
Anthropic is really going on the defense against prompt injection uh for browser agents, um, you know, with uh making sure uh that virtual machines are isolated and have minimal privileges, uh, and going with domain allowed list versus blacklists, right?
00:25:22.480 --> 00:25:24.880
So everyone has a little bit of a different uh approach here.
00:25:24.960 --> 00:25:34.559
You know, Google, as an example, there, uh Project Mariner, uh browser agents run in virtual machines as a safety uh to isolate them from anything else they can kind of get their hands on.
00:25:34.640 --> 00:25:35.839
And then Microsoft, right?
00:25:35.920 --> 00:25:39.839
There's millions of ways that Microsoft is doing it and millions of ways that all these companies are doing it.
00:25:39.920 --> 00:25:44.000
I just kind of picked out uh you know different aspects to illustrate their different approach.
00:25:44.160 --> 00:25:49.519
You know, Microsoft, as an example with their co-pilot studio, I mean the governance is everywhere, right?
00:25:49.599 --> 00:25:56.000
They have the Sentinel monitoring, the uh the purview logging, you know, the agent uh intra ID, uh, right.
00:25:56.079 --> 00:25:59.440
So Microsoft, it is very enterprise, Google is as well.
00:25:59.519 --> 00:26:04.640
I was just giving an example of you know how their project mayor kind of runs in an isolated virtual machine.
00:26:04.799 --> 00:26:18.319
Uh because essentially, I think from the people I've talked to, um, at least uh three of those four companies, um they know risk is part of it, right?
00:26:18.640 --> 00:26:21.920
Like there's like any lab, right?
00:26:22.079 --> 00:26:25.599
They call them frontier, you know, or AI labs for a reason.
00:26:25.759 --> 00:26:31.119
You run experiments, and part of it is well, you know that things need to go wrong, right?
00:26:31.200 --> 00:26:39.200
So these companies are intentionally trying to get all the risk and all the security nightmares and all the sprawl they can, right?
00:26:39.279 --> 00:26:45.119
So then when they set a new model out in the wild, when they're done with the red teaming, um, you know, they hope that they have a good idea.
00:26:45.279 --> 00:27:02.640
So the labs are working on this, but here's the reality I think there's a pressure to allocate more resources to the development of models versus research and security, right?
00:27:03.359 --> 00:27:06.319
I don't have that on you know, privileged and information.
00:27:06.400 --> 00:27:12.160
That's just from you know, talking to a lot of people and reading a lot and just the I think the reality of the world that we live in.
00:27:12.799 --> 00:27:19.279
And that's why AI sprawl is going to be hard to miss for most business leaders.
00:27:19.440 --> 00:27:22.480
Um, because everything's unclear, right?
00:27:22.559 --> 00:27:24.079
The tool, like, do you know?
00:27:24.240 --> 00:27:26.880
Think of the the the AI model you use most.
00:27:27.599 --> 00:27:28.880
Do you know the tools it has?
00:27:29.519 --> 00:27:29.759
Right.
00:27:29.920 --> 00:27:36.640
Even if you think of, you know, Chat GPT, Gemini, Claude, uh, do you know what tools the base models have?
00:27:37.279 --> 00:27:38.240
Probably not.
00:27:38.960 --> 00:27:40.559
You'd be surprised, right?
00:27:40.720 --> 00:27:46.160
Like, I'm one of those weird guys that that reads uh change logs um and model cards.
00:27:46.319 --> 00:27:48.400
Most people have no clue, right?
00:27:48.640 --> 00:28:00.960
And this is much worse than classic shadow IT because agents can act across systems, not just you know, within the confines of uh folders and files, which are very structured.
00:28:01.200 --> 00:28:08.160
The whole kind of um, it's kind of like how people talk about hallucinations, right?
00:28:08.319 --> 00:28:11.200
Um, oh, you know, they're not a bug, they're a feature, right?
00:28:11.359 --> 00:28:15.680
What makes agents great is also the built-in nightmare.
00:28:15.759 --> 00:28:20.960
So if you want to have the good, you have to recognize the bad that they can do.
00:28:21.200 --> 00:28:22.000
Um, right.
00:28:22.079 --> 00:28:23.759
And that's because that's how they're made.
00:28:23.920 --> 00:28:29.200
They are made to go build their own path, they are made to blaze their own trail, right?
00:28:29.359 --> 00:28:35.839
So by their very definition, agents aren't necessarily always good at staying within guardrails, right?
00:28:35.920 --> 00:28:44.480
Because if they think that the destination that has been given to them, um, at times, right, depends on the model to set up, all that, right?
00:28:44.640 --> 00:28:53.599
But it's very common for an agent to hop over guardrails in order to accomplish a task, not knowing that that's a bad thing, right?
00:28:54.000 --> 00:29:05.839
And every new skill that you add that you give to uh an agent or a model, every connector, every workflow, uh, all that does is it expands the attack surface for everyone.
00:29:06.240 --> 00:29:12.079
And that's another realization that I think most people are going to struggle to grasp, right?
00:29:12.160 --> 00:29:18.240
Because you think that you're expanding the amount of work that you can get done.
00:29:18.480 --> 00:29:19.440
And are you?
00:29:19.759 --> 00:29:20.240
Sure.
00:29:20.400 --> 00:29:26.319
But I think you almost have to think of it from like a, I don't know, like an old school like military perspective, right?
00:29:26.400 --> 00:29:27.680
It's your land, right?
00:29:27.839 --> 00:29:29.680
And hey, we're gonna go discover new land.
00:29:29.759 --> 00:29:32.880
We're gonna go out and, you know, win new deals, right?
00:29:33.039 --> 00:29:34.559
So we're gonna expand our land.
00:29:34.640 --> 00:29:37.279
I don't know if any of you have ever played like, you know, Risk.
00:29:37.759 --> 00:29:38.960
Um, I used to.
00:29:39.119 --> 00:29:40.319
I wish I still had time, right?
00:29:40.400 --> 00:29:44.799
It's one of those games, you know, it's like, oh, okay, this is good to make sure my brain still works.
00:29:44.880 --> 00:29:46.400
But it's kind of like that, right?
00:29:46.559 --> 00:29:51.279
Um, your agents, without you knowing, though, are acquiring new territories.
00:29:51.440 --> 00:29:52.240
And that's great.
00:29:52.319 --> 00:29:55.680
And you think, oh my gosh, look at these new skills and capabilities, right?
00:29:55.759 --> 00:29:58.400
It's like, oh, it's like having a thousand new employees.
00:29:58.559 --> 00:29:59.200
Okay, great.
00:29:59.359 --> 00:30:08.799
But your surface area for risk, attacks, and sprawl is also multiplying at the same time as your capability.
00:30:08.960 --> 00:30:10.319
But you're not focused on that.
00:30:10.480 --> 00:30:13.759
You're focused on, oh my gosh, now all of a sudden I have a designer.
00:30:13.839 --> 00:30:15.839
Now all of a sudden I have a data analyst.
00:30:16.000 --> 00:30:20.079
Okay, well, what's that data analyst doing with your data?
00:30:20.400 --> 00:30:21.920
Are you going to check?
00:30:22.240 --> 00:30:22.559
Right?
00:30:22.799 --> 00:30:25.440
Is it uh using an in open source, right?
00:30:25.519 --> 00:30:28.319
This is one thing I was kind of blown away by the other day.
00:30:28.720 --> 00:30:33.599
Uh gave, gave uh, you know, it wasn't anything bad, right?
00:30:34.000 --> 00:30:35.839
Gave a model a very hard task.
00:30:36.000 --> 00:30:45.519
This is one of those that uh overnight um I came back and I realized, oh, it downloaded another large language model locally to do the task, right?
00:30:45.599 --> 00:30:47.359
It did it locally, uh, right.
00:30:47.440 --> 00:30:52.319
So uh kept all the data uh kind of on on-prem on my local machine.
00:30:52.640 --> 00:30:59.200
But what happens if in the future other AI models are just gonna go use other AI models, but on the web, right?
00:30:59.359 --> 00:31:00.079
They can.
00:31:00.319 --> 00:31:02.319
What if they do that without you telling?
00:31:02.480 --> 00:31:06.240
Or what if they go use a um a site that's unsecure?
00:31:06.319 --> 00:31:13.200
And and that, and this is why I think, you know, open claw, as amazing as it is, this is why it's also scary, right?
00:31:13.599 --> 00:31:26.000
Because, you know, these you know, uh autonomous open source agents, the open source, if I'm being honest, and I'm sure, you know, uh I'm gonna catch some flack from this from some OG open source people, right?
00:31:26.160 --> 00:31:29.200
Um, OG open source, I'm not talking about you.
00:31:29.359 --> 00:31:35.039
Today's open source, it's taken a weird turn over the past uh three to six months, right?
00:31:35.200 --> 00:31:42.160
It's almost become this, you know, crypto-infused, wild, wild web point four west, right?
00:31:42.240 --> 00:31:45.279
It's not like traditional open source anymore.
00:31:45.359 --> 00:31:53.119
So when we talk about open source AI agents, right, so many of what's so much of what's out there is risk, right?
00:31:53.359 --> 00:31:58.559
It is, and it's unknown and it can be scary if you don't know exactly what you're getting into.
00:31:59.680 --> 00:32:04.400
And I think this is why small agent pilots are becoming huge risk, right?
00:32:04.559 --> 00:32:08.880
Because when you build your own agents without centralized visibility or governance, that's where you get in trouble.
00:32:08.960 --> 00:32:12.319
So many people and they're using their own agents, it's for them, right?
00:32:12.559 --> 00:32:18.960
But it can be go going and doing work for the whole department, the whole company, or accessing data from the whole company.
00:32:19.039 --> 00:32:26.640
But if one person has eyes on it, and if there's not central um organization, that's where you run into something.
00:32:27.519 --> 00:32:29.759
So here's the playbook, right?
00:32:30.079 --> 00:32:33.039
If you're feeling a little scared, you know what?
00:32:33.119 --> 00:32:34.319
Maybe that was my intent.
00:32:34.480 --> 00:32:44.000
I got y'all pumped up and um, you know, excited with yesterday's show talking about, you know, the new capabilities of agents, but you gotta get it in check, right?
00:32:44.160 --> 00:32:47.279
You have to balance that optimism with a little bit of realism.
00:32:47.359 --> 00:32:52.960
But here is the Monday morning playbook for how to deal with the risk, the attacks, and the sprawl.
00:32:53.039 --> 00:32:58.400
Uh so like what we talked about yesterday, touched on it briefly, but you need to start with bounded autonomy, right?
00:32:58.640 --> 00:33:04.000
When you think of autonomous agents, it's not uh, you know, zero to a hundred.
00:33:04.079 --> 00:33:05.200
It's bounded autonomy.
00:33:05.279 --> 00:33:13.920
What that means is you start with suggesting, then propose, then approve, then limited um execution, right?
00:33:14.079 --> 00:33:16.480
So many people go to full execution first.
00:33:16.960 --> 00:33:18.720
It's baby steps, right?
00:33:19.039 --> 00:33:23.359
Uh a human being doesn't go from womb to sprint, right?
00:33:23.440 --> 00:33:33.359
You go from womb to you know being being on your stomach to sitting, to crawling, to walking, to right?
00:33:33.759 --> 00:33:35.359
Your agents have to go the same way.
00:33:35.440 --> 00:33:41.759
Even if you know out of the box that agent can sprint, you're not prepared for that agent to sprint.
00:33:41.920 --> 00:33:43.759
Part of it is for the humans, right?
00:33:43.920 --> 00:33:47.920
We think it's more for um understanding the the agent's abilities.
00:33:48.000 --> 00:33:49.680
It's not, right?
00:33:50.079 --> 00:33:53.200
This is to make sure that you don't have a bunch of lazy humans in the loop.
00:33:53.279 --> 00:33:57.599
And instead, you have uh proactive expert-driven loops.
00:33:57.759 --> 00:34:00.720
Um so start with least privilege by default.
00:34:00.880 --> 00:34:05.599
Read only first, write access only for narrow defined tasks, right?
00:34:05.759 --> 00:34:10.800
Especially in enterprise environments, you don't want to send out a bunch of read-only general agents.
00:34:11.039 --> 00:34:14.960
Start with read only first, observe, uh, improve your loop.
00:34:15.039 --> 00:34:21.440
Then you go to um limiting, uh, limited execution.
00:34:21.760 --> 00:34:28.159
Um, and then you can finally get it to um writing for narrow tasks.
00:34:28.320 --> 00:34:38.159
Uh, and then last in your Monday morning playbook is you need to require human approvals for irreversible actions like sends, deletes, purchases, and permission changes.
00:34:38.400 --> 00:34:40.800
These are all things agents can do, right?
00:34:40.960 --> 00:34:42.559
Agents can literally do anything.
00:34:42.719 --> 00:34:52.559
Agents, uh, right, and this is going to get even harder to manage when we start to see true agentic commerce, right?
00:34:52.639 --> 00:34:59.519
And I'm not just saying, oh, you know, my agent is gonna go buy something, you know, from uh the Amazon agent.
00:34:59.679 --> 00:35:01.039
That's not what I'm saying.
00:35:01.119 --> 00:35:01.599
Uh, right.
00:35:01.679 --> 00:35:08.960
There's going to be agent uh bartering, there's going to be agent um, you know, liaisons, right?
00:35:09.039 --> 00:35:13.760
So uh you have to work on human approvals now, right?
00:35:13.920 --> 00:35:29.280
It's almost like you have to understand that agents are capable of you know a 10 and you have to give them a two because we as humans need to first learn to adapt our behavior um before just giving in to agents, right?
00:35:30.000 --> 00:35:33.360
And then you need to build governance before you scale.
00:35:33.599 --> 00:35:36.079
And I think that's like just what I was getting at.
00:35:36.239 --> 00:35:36.480
Right.
00:35:36.719 --> 00:35:40.800
We get all excited for what we know it can do.
00:35:40.880 --> 00:35:43.360
And then we're like, wait, we got to do this boring stuff first?
00:35:43.599 --> 00:35:44.400
Absolutely.
00:35:44.639 --> 00:35:52.000
Because if you don't do the boring stuff first, if you don't go through the sitting on your tummy, right?
00:35:52.800 --> 00:36:00.800
Sitting up, crawling, walking, then you can't expect to understand uh the path of the sprinting agent.
00:36:00.960 --> 00:36:04.719
And every agent run needs a decision trace that you can expect after the fact.
00:36:04.800 --> 00:36:11.039
So you have to be able to log all your tool calls, capture your decision traces, and monitor for abnormal action patterns.
00:36:11.199 --> 00:36:18.719
This is one of the reasons why I said, I think we're gonna see in 2026, uh, it's gonna be extremely common to have agent op teams, right?
00:36:18.800 --> 00:36:23.280
In the same reason, or sorry, in the same vein, you know, dev op teams are very common, right?
00:36:23.360 --> 00:36:26.480
We're gonna have agent op teams for this exact reason.
00:36:26.960 --> 00:36:32.880
So here's what to watch for for the rest of 2026 uh when it comes to agent risk security and sprawl.
00:36:33.119 --> 00:36:36.400
Uh browser agents are gonna become a mainstream risk surface.
00:36:36.480 --> 00:36:37.360
I think we know that.
00:36:37.440 --> 00:36:40.400
Uh, I think we're gonna see a major open claw type.
00:36:40.480 --> 00:36:45.519
I'm not saying open claw themselves, but I think we're gonna see a major uh open source agent crash.
00:36:45.679 --> 00:36:48.559
I talked about that in the 2026 AI prediction and roadmap series.
00:36:48.719 --> 00:36:50.320
I actually think it's gonna come from, right?
00:36:50.400 --> 00:36:59.440
There's this new trend, like trend over the last like week or two, where everyone is just sending pointing their agents to a uh a website, right?
00:36:59.599 --> 00:37:12.159
When they want their agent to do something, they're not even saying, all right, let me uh let me, the human, understand this, let me write it out, let me apply this to my business logic, let me apply this with my garden.
00:37:12.239 --> 00:37:12.480
Nope.
00:37:12.559 --> 00:37:17.039
They're just saying pointing their agent, hey agent, here's this video, go watch it and do it.
00:37:17.119 --> 00:37:22.079
Hey, agent, here's this website with a whole bunch of cool stuff, just go do it.
00:37:22.320 --> 00:37:26.880
You know, there's this whole concept now going around of, you know, point your agent to a URL.
00:37:27.599 --> 00:37:29.280
Okay, have fun with that.
00:37:29.679 --> 00:37:31.360
Don't do that, right?
00:37:31.599 --> 00:37:32.400
Don't, right?
00:37:32.480 --> 00:37:35.440
Especially if that agent has its own dedicated machine, right?
00:37:35.519 --> 00:37:38.320
And I'm saying if you're tinkering around, right?
00:37:38.400 --> 00:37:41.679
It's it's it's your personal computer, you're the business owner.
00:37:41.760 --> 00:37:43.360
If you want to take that risk, that's fine, right?
00:37:43.440 --> 00:37:45.840
Especially if you're you're you're tinkering around, that's fine.
00:37:45.920 --> 00:37:47.679
But don't do that in an enterprise.
00:37:47.840 --> 00:37:52.960
And uh unfortunately, people are doing that, and that's a bad thing because guess what?
00:37:53.119 --> 00:37:57.599
A lot of those sites, quote unquote, oh, you know, someone puts up a huge tutorial.
00:37:57.679 --> 00:37:59.119
Uh okay.
00:38:00.079 --> 00:38:01.920
They're putting it up on their personal website.
00:38:02.000 --> 00:38:06.239
You know how easy it is to hack, uh, to fish any of these websites?
00:38:06.480 --> 00:38:08.239
Okay, you see, let me give you an example.
00:38:08.400 --> 00:38:09.679
Something goes viral on Twitter.
00:38:09.840 --> 00:38:16.719
It's some, you know, uh literally, some 20-year-old kid, you know, put a great uh, you know, guide up on the website.
00:38:16.880 --> 00:38:17.119
Cool.
00:38:17.199 --> 00:38:20.559
And, you know, everyone's just like, hey, agent, go read that, go read that.
00:38:20.639 --> 00:38:25.039
And then you have millions of agents going to just read this, you know, this nice kid, put up a cool guide.
00:38:25.280 --> 00:38:26.960
Okay, guess what's gonna happen?
00:38:27.119 --> 00:38:32.639
Someone's gonna see that, they're gonna hack that site and they're gonna inject malicious code in there that no one's gonna see.
00:38:32.719 --> 00:38:33.679
That's what's gonna happen.
00:38:33.920 --> 00:38:38.719
All right, next, the agent skill marketplace is going to expand uh the supply chain risk.
00:38:38.800 --> 00:38:40.400
So you inherit what you plug in.
00:38:40.480 --> 00:38:44.400
And we've already seen that with uh some skills and plugins for agents.
00:38:44.480 --> 00:38:48.559
Uh, some of the top ones were found to have uh you know malicious code in them.
00:38:48.639 --> 00:38:55.119
Uh, you need to uh we're gonna see identity and permissions becoming a board level compliance requirement, not a nice to have.
00:38:55.280 --> 00:39:01.039
And last but not least, the winners are gonna treat agents like production software, not side experiments.
00:39:01.199 --> 00:39:02.159
All right, that's it.
00:39:02.239 --> 00:39:05.360
That's a wrap volume nine of the start here series.
00:39:05.599 --> 00:39:06.480
I hope this is helpful.
00:39:06.559 --> 00:39:27.679
I hope this was helpful and I hope you understand a little bit more the risk security and the sprawl uh that you need to be aware of and how this having AI that acts, having AI, right, that has a smart brain, it's it's proactive, it has tools, and it has arms.
00:39:28.559 --> 00:39:30.480
That changes everything, right?
00:39:30.559 --> 00:39:42.559
This perfect storm that has happened in the last 30 days, we've seen, I've said this before, we've seen more both on the opportunity and the risk side.
00:39:42.719 --> 00:39:47.920
We've seen more in the last 30 days than we've seen in the last probably two years.
00:39:48.159 --> 00:39:48.719
All right.
00:39:48.960 --> 00:39:54.320
So I wanted to take a moment for the start here series to address this.
00:39:54.480 --> 00:40:03.920
And I hope this helps you not only understand the risk, but if you want to take advantage of the opportunities, you can't do it without knowing the risk.
00:40:04.079 --> 00:40:04.719
So now you do.
00:40:04.880 --> 00:40:06.480
All right, so I hope this was helpful.
00:40:06.639 --> 00:40:12.719
So if so, there's already eight other uh in our start here series that you should go check out.
00:40:12.800 --> 00:40:15.920
So, like I said, please go to starthearseries.com.
00:40:16.000 --> 00:40:19.360
That is gonna give you free access to our inner circle community.
00:40:19.519 --> 00:40:34.079
Once you sign up, you're gonna be uh just uh straight up loving it, hopefully, uh, in our start here series space and connecting uh with now more than a thousand people uh in our uh inner circle community.
00:40:34.239 --> 00:40:35.360
So thank you for tuning in.
00:40:35.519 --> 00:40:38.320
Hope to see you back for more everyday AI.
00:40:38.400 --> 00:40:39.360
Thanks, y'all.
00:00:01.040 --> 00:00:03.680
There's always been AI risk.
00:00:03.919 --> 00:00:18.960
But in the early days of large language models and chatbots, that risk was like Bill getting something wrong in the blog post or Deborah putting up a hallucinated stat in the onboarding guide that was littered with M dashes and delves.
00:00:19.600 --> 00:00:26.239
But AI Risk Today is a legit different ball game than risk was three and a half years ago.
00:00:26.399 --> 00:00:31.519
I mean, heck, AI Risk today is unrecognizable from what it was three and a half months ago.
00:00:31.679 --> 00:00:33.439
And that's not an exaggeration.
00:00:33.679 --> 00:00:41.119
Because after hearing for like five years that we're six months away from real AI agents, well, it finally happened.
00:00:41.280 --> 00:00:50.799
And it was actually this perfect storm of multiple events that led to an unexpected business scenario that the corporate world now faces today.
00:00:50.960 --> 00:00:58.640
You either get on board with AI agents quickly or get left behind, but do it too quickly, and you could go under.
00:00:59.200 --> 00:01:02.640
The risk, the security, and the sprawl are real.
00:01:02.799 --> 00:01:07.840
And so we're gonna tackle it all today on Everyday AI.
00:01:08.000 --> 00:01:10.000
The start here, series edition.
00:01:10.239 --> 00:01:12.079
All right, well, welcome to Everyday AI.
00:01:12.159 --> 00:01:13.280
My name's Jordan Wilson.
00:01:13.359 --> 00:01:14.959
And if you're new here, this thing's for you.
00:01:15.040 --> 00:01:23.840
It's your daily live stream podcast and free daily newsletter helping everyday business leaders like you and me keep up with all the news, like a thousand new agents a day.
00:01:23.920 --> 00:01:24.400
What do we do?
00:01:24.480 --> 00:01:25.040
What do we try?
00:01:25.120 --> 00:01:30.159
Well, tune in, I tell you, and uh help you make the right decisions to grow your company and your career.
00:01:30.319 --> 00:01:36.159
So after 700 plus episodes, I realized I couldn't answer the most common question that people had for me.
00:01:36.239 --> 00:01:37.920
Like Jordan, you have a lot of episodes.
00:01:38.000 --> 00:01:38.719
Where do I start?
00:01:38.879 --> 00:01:41.920
Well, that's why I created the Start Here series.
00:01:42.000 --> 00:01:49.439
So the Start Here series is the essential podcast series to both learn the AI basics and to double down on your knowledge.
00:01:49.599 --> 00:01:59.439
Uh so this is volume nine, and you can go listen to all of the uh episodes in our Start Here series if you just go to starthearseries.com.
00:01:59.519 --> 00:02:08.159
So that will give you uh free access to our inner circle community, and it'll put you right in the Start Here Series space.
00:02:08.319 --> 00:02:15.039
So you can go uh listen to all of them, watch them, read about them all in one place, interact with others who are doing the same.
00:02:15.199 --> 00:02:19.680
All right, so uh one other thing that you need to do before we get started.
00:02:19.919 --> 00:02:23.919
My gosh, y'all, you you have to go listen to these.
00:02:24.000 --> 00:02:26.960
Uh, episode 712 and 713.
00:02:27.120 --> 00:02:30.159
That is our 2026 AI prediction and roadmap series.
00:02:30.240 --> 00:02:38.800
That's like a culmination of a thousand hours of work over the years uh or over the past year uh to give you guys the blueprint for 2026.
00:02:38.879 --> 00:02:40.879
So make sure you go listen to those.
00:02:41.120 --> 00:02:41.520
All right.
00:02:41.680 --> 00:02:46.080
Uh, if this sounds kind of like our last episode in the Start Here series, not exactly.
00:02:46.159 --> 00:02:48.479
Uh so make sure you go listen to that one if you didn't already.
00:02:48.560 --> 00:02:55.280
So this is more of just the state of AI agents, where they are, what they are, how we should use them, should we use them.
00:02:55.439 --> 00:02:58.639
So make sure you go listen to uh volume eight from yesterday.
00:02:58.800 --> 00:03:00.479
That's episode 717.
00:03:00.879 --> 00:03:07.759
But today we're here to talk about the other side, the risk, the security, and the sprawl that's changing everything.
00:03:08.080 --> 00:03:14.879
So here's what we're gonna be covering on today's show and well, why it urgently matters.
00:03:15.199 --> 00:03:20.319
Because AI models didn't just get smarter, they got hands, right?
00:03:20.400 --> 00:03:29.199
Uh the risk model changed when AI moved from generating text like it was three and a half years ago, to now it's taking real actions.
00:03:29.280 --> 00:03:32.560
And a lot of times actions we're not aware of, and that's the scary part.
00:03:32.719 --> 00:03:43.919
And an agent connected to your email and calendar or you know, your company's data can act fast, confident, and wrong in the same way that AI models hallucinated three years ago.
00:03:44.080 --> 00:03:45.680
Well, they can still hallucinate now.
00:03:45.840 --> 00:03:48.560
And this isn't a future concern, this is happening right now.
00:03:48.719 --> 00:03:55.919
So on today's show, uh, we're gonna go over a simple mental model for why agent risk is fundamentally different from chatbot risk.
00:03:56.000 --> 00:04:02.479
Uh, we're gonna talk about what OpenAI, Google, Anthropic, and Microsoft are actually building to address this risk.
00:04:02.639 --> 00:04:11.599
And I'm gonna give you a practical Monday morning playbook that you can start using this week to address the risk security and sprawl.
00:04:11.919 --> 00:04:12.800
Sound good?
00:04:13.120 --> 00:04:15.120
Yeah, sounds good to me.
00:04:15.199 --> 00:04:17.839
I'm I'm excited and I wrote all this.
00:04:18.000 --> 00:04:21.279
All right, so let's get a quick little catch up here.
00:04:21.439 --> 00:04:26.480
Um, so probably if you're listening to this show, AI is not new to you, right?
00:04:26.639 --> 00:04:30.560
But um, let me just give everyone the the the briefer here, right?
00:04:30.720 --> 00:04:39.279
So essentially from 2022 to mid-2023, early 2024, large language models were largely text systems, right?
00:04:39.439 --> 00:04:47.040
With the the risk was just limited to misinformation and data leaks and you know, using hallucinations and looking foolish.
00:04:47.199 --> 00:04:51.279
Uh, but that started to change, I'd say, in the mid-2025.
00:04:51.439 --> 00:04:58.160
Uh, so that's when the models started getting well exponentially more capable and agentic by default.
00:04:58.319 --> 00:05:01.120
That's the thing that people don't understand.
00:05:01.279 --> 00:05:01.680
Uh, right.
00:05:01.759 --> 00:05:11.759
I've I've actually had two great conversations with uh two really smart minds, uh, kind of the head of um agents at Cloudflare and then uh the head of Microsoft Research.
00:05:11.839 --> 00:05:15.600
And I learned a lot by talking to them both, you know, before and after the show as well.
00:05:15.759 --> 00:05:21.839
Um but one thing that kind of came true is well, no one really knows what constitutes an agent and what doesn't.
00:05:22.000 --> 00:05:28.720
But I think that we can all agree that even today's large language models, they're agentic by nature, right?
00:05:28.879 --> 00:05:33.439
They, well, a lot of us are giving them access to all of our data.
00:05:33.600 --> 00:05:39.279
Um, and not in all cases do they have right access, but in many cases, they do.
00:05:39.519 --> 00:05:47.600
Um, so when these models can think and act on their own and um, you know, spin up a virtual environment and a terminal and access your computer, right?
00:05:47.759 --> 00:05:48.639
I keep saying this.
00:05:48.800 --> 00:05:53.519
Even right now, I have multiple agents running on my computer, right?
00:05:53.680 --> 00:05:56.480
I have codecs and clawed code going right now.
00:05:56.560 --> 00:05:59.920
I'll probably spin up uh anti-gravity uh later.
00:06:00.319 --> 00:06:06.240
I always have agents running and they have access, and I don't necessarily know what they're doing uh, you know, in between.
00:06:06.319 --> 00:06:07.839
I always go back and look when they're done.
00:06:08.079 --> 00:06:14.319
But that's what's really changed um in mid uh, you know, 2025.
00:06:14.480 --> 00:06:34.319
But here we are in early 2026, and this is where uh now we're gonna start talking about and actually putting into practice all those buzzwords, uh, right, that people just you know started chatting about in like 2024 to you know sound super smart, like oh, governance and audit logs and you know, uh isolation uh protocols, right?
00:06:34.399 --> 00:06:35.680
Like all those well, okay.
00:06:35.759 --> 00:06:40.959
Well, good thing that everyone was talking about it, uh, because you know, we were trying to get our our buzzword bingo.
00:06:41.040 --> 00:06:42.879
Um well, now we actually need it, right?
00:06:42.959 --> 00:06:50.160
Now it's uh time to talk about uh ethics and govern governance and guardrails when it comes to uh agentic capabilities.
00:06:55.279 --> 00:06:58.480
I like to put it like this to keep it very simple.
00:06:59.759 --> 00:07:09.839
Um 2022 and before, I'd say that AI was a dumb stationary brain, but it was a brain, right?
00:07:09.920 --> 00:07:12.079
So everyone was blown away, like, oh my gosh, it can think.
00:07:12.240 --> 00:07:13.120
Well, it was dumb.
00:07:13.279 --> 00:07:17.279
It was stationary, didn't move, couldn't really think ahead, right?
00:07:17.439 --> 00:07:22.480
And in 2023, well, it became a dumb stationary brain with tools, right?
00:07:22.560 --> 00:07:29.600
That's when uh, you know, the early version of uh ChatGPT plus uh, you know, GPT 4, it had tools, right?
00:07:29.680 --> 00:07:31.439
It could go on the internet as an example.
00:07:31.519 --> 00:07:37.360
So that's when it uh really started to open up what it could do, or at least the information that it could access.
00:07:37.600 --> 00:07:44.480
In 2024, uh, well, it was still a stationary brain with tools, but it went from a dumb brain to a smart brain.
00:07:44.560 --> 00:07:51.759
I would say in 2024 was the first time that we actually had smart models, uh, because at the end of 2024, that's when we got reasoning models.
00:07:51.839 --> 00:08:02.399
Uh and then in 2025, I think we still had smart brains with tools, but the difference now, instead of it being a stationary brain, it was a proactive brain in 2025.
00:08:02.639 --> 00:08:10.879
It could go out, at least, especially at the end of 2025, the second and third quarter, it can make moves on its own, right?
00:08:11.439 --> 00:08:12.240
Proactively.
00:08:12.800 --> 00:08:18.720
It could schedule things and or just you know, they can start acting over long periods.
00:08:18.959 --> 00:08:28.319
And then what brings it to 2026 is well, now we have that smart, proactive brain with tools and arms, right?
00:08:28.720 --> 00:08:36.240
So the tools are cool, but when you have arms, you can actually use them in a real way.
00:08:36.320 --> 00:08:44.559
And I think that's where you know agents now have teeth when they are autonomous, uh proactive, and smart.
00:08:45.279 --> 00:08:47.919
So here's why agent risk I think feels different.
00:08:48.159 --> 00:08:53.600
Um, because you know, like I said, with the chatbots, there was always risk, but was it really?
00:08:54.000 --> 00:09:00.159
I mean, yeah, worst thing you can do is you get in trouble putting out something hallucinated and you look foolish, right?
00:09:00.399 --> 00:09:02.000
Uh does your company go under?
00:09:02.159 --> 00:09:03.039
Probably not.
00:09:03.279 --> 00:09:03.840
Uh right.
00:09:03.919 --> 00:09:06.480
Do you expose every single dark sea, right?
00:09:06.639 --> 00:09:07.600
It's not that bad.
00:09:07.840 --> 00:09:09.440
I mean, it is, but it's not.
00:09:09.600 --> 00:09:16.639
But that the the new agent layer is just a whole new type of risk that we're not even ready for.
00:09:16.799 --> 00:09:20.240
If I'm being honest, if you listen to the show, I say this a lot.
00:09:20.399 --> 00:09:22.720
I'm like, this year is going to be scary.
00:09:23.279 --> 00:09:27.200
People are not ready, businesses are not ready, and I mean that, right?
00:09:27.440 --> 00:09:32.399
That kind of business predicament that I talked about in the opening of the show there, that is real.
00:09:33.120 --> 00:09:39.679
If you don't run to use AI agents this year, you're toast.
00:09:40.399 --> 00:09:41.279
You are literally toast.
00:09:41.840 --> 00:09:50.159
I don't care if you're a uh small business or a$20 billion revenue business, you're toast.
00:09:50.480 --> 00:09:50.799
All right.
00:09:51.679 --> 00:09:54.879
But if you sprint too quickly, you can go under, right?
00:09:55.279 --> 00:10:02.960
Because you could make devastating mistakes that that would be uh highly improbable to recover from, right?
00:10:03.120 --> 00:10:09.440
Because AI agents can do things uh much worse than even the worst human, right?
00:10:09.600 --> 00:10:12.320
Have you ever heard stories or maybe you've experienced this, right?
00:10:12.480 --> 00:10:16.000
Kind of a rogue employee happens too often, right?
00:10:16.080 --> 00:10:21.840
Someone's someone's bitter maybe about getting fired or not getting a promotion, and they do something absolutely crazy, right?
00:10:21.919 --> 00:10:26.159
Maybe expose all the company secrets or you know, release some files.
00:10:26.320 --> 00:10:26.879
I don't know.
00:10:27.600 --> 00:10:29.440
Okay, that's a human, right?
00:10:29.600 --> 00:10:32.559
That's a single human, and you can see that human.
00:10:32.639 --> 00:10:34.639
You have eyes on that human, right?
00:10:34.879 --> 00:10:37.200
Bill and IT is watching that human.
00:10:38.080 --> 00:10:39.039
Agents are different.
00:10:40.000 --> 00:10:46.240
Agents move a hundred, a thousand times faster than that one person, but you can't always see agents.
00:10:46.639 --> 00:10:47.200
And guess what?
00:10:47.279 --> 00:10:49.440
That one disgruntled employee, that's one.
00:10:50.240 --> 00:10:51.200
This isn't a video game.
00:10:51.279 --> 00:10:52.480
You can't respond 10 times.
00:10:52.639 --> 00:10:53.360
Agents can.
00:10:54.000 --> 00:11:00.240
Agents can spawn sub-agents like that, and those sub-agents can spawn like that, right?
00:11:00.399 --> 00:11:08.720
So think of in the way like a virus might spread across your computer or across the human body, uh, right, and replicate and duplicate.
00:11:08.879 --> 00:11:10.480
It's the same thing with AI agents.
00:11:10.639 --> 00:11:12.080
A rogue employee can't do that.
00:11:12.159 --> 00:11:14.639
And that's why this risk is very different.
00:11:14.720 --> 00:11:16.000
It is very real.
00:11:16.559 --> 00:11:25.919
So I think that we spend too much time uh thinking about the positives and the optimistic side of AI agents, which is great, right?
00:11:26.080 --> 00:11:33.440
Yes, oh now all of a sudden I have, you know, 320 uh AI agents, you know, doing all my work for me around the clock.
00:11:33.519 --> 00:11:34.320
Oh, that's cool.
00:11:34.559 --> 00:11:34.960
All right.
00:11:35.200 --> 00:11:38.240
But what about the risk, the security, and the sprawl?
00:11:39.440 --> 00:11:46.080
And more teams are experimenting now than ever before because no one's got it all figured out.
00:11:46.240 --> 00:11:52.240
And that means more sprawl uh and more exposure, uh, especially if you don't have guardrails up.
00:11:52.399 --> 00:11:56.080
So here's essentially the three surfaces where agent risk actually lives.
00:11:56.240 --> 00:11:57.360
Number one is the input.
00:11:57.519 --> 00:12:03.840
That's kind of that untrusted content that can contain hidden instructions, agents you know, treat like real commands.
00:12:04.000 --> 00:12:07.759
All right, that I don't think is going to be as big of a deal, right?
00:12:07.919 --> 00:12:11.919
Your inputs, don't get me wrong, things can go wrong, right?
00:12:12.080 --> 00:12:14.240
People are blindly copying and pasting.
00:12:14.320 --> 00:12:19.679
So you can copy and paste prompt injections, but prompt injections are the big thing, and we're gonna talk about that here in a second.
00:12:19.759 --> 00:12:22.559
But still, inputs, they can be poisoned, right?
00:12:22.639 --> 00:12:24.559
There's there's things that you might not know.
00:12:24.720 --> 00:12:33.679
And in the same way, a rogue human can create a lot of rogue AI agents that can create a lot of risk and a lot of sprawl that's gonna be uncontrollable.
00:12:33.759 --> 00:12:42.000
So, you know, inputs, I think you really only have to worry about it, you know, if someone really doesn't know what they're doing or if someone is trying to be malicious, which again, those things are gonna come up.
00:12:42.159 --> 00:12:46.960
But think of that one bad employee, what they can now do if they know agentic AI, right?
00:12:47.120 --> 00:12:47.519
Yeah.
00:12:47.679 --> 00:12:58.480
Uh talk about malware, spy spyware, uh, right, ransomware that we, you know, these companies that have paid, you know, millions of dollars, hundreds, tens of millions of dollars for ransomware.
00:12:58.559 --> 00:13:00.240
It's gonna be way worse with agents.
00:13:00.320 --> 00:13:02.159
Uh, so the second layer is tools.
00:13:02.240 --> 00:13:09.200
And I think this is where it starts to get a little dicey in terms of, well, the capabilities are wild, right?
00:13:09.360 --> 00:13:15.600
So every permission and connector that you add expands that blast radius essentially when something goes wrong.
00:13:15.759 --> 00:13:19.519
In the same way that I walked you guys through the, oh, it's a dumb stationary brain.
00:13:19.679 --> 00:13:23.440
But hey, once that brain gets tools, okay, now it can start doing a lot of things.
00:13:23.600 --> 00:13:26.799
And the same thing on the agent gone wrong side, right?
00:13:26.960 --> 00:13:33.679
If it was a dumb stationary uh agent with no tools and no arms, it's like, all right, well, have fun, buddy.
00:13:33.759 --> 00:13:34.000
Right?
00:13:34.080 --> 00:13:36.559
You're in a glass case of emotion just shaking up.
00:13:36.720 --> 00:13:39.360
Uh when they have tools, that's where things go wrong, right?
00:13:39.519 --> 00:13:43.679
When they have uh access to your terminal, right?
00:13:43.759 --> 00:13:45.919
To your computer terminal, that's where things can go wrong.
00:13:46.080 --> 00:13:49.279
When they can run code on a machine, that's where things can go very wrong.
00:13:49.440 --> 00:13:53.519
And then last but not least, and this is the big one, this is actions.
00:13:53.759 --> 00:14:04.720
And this is the biggest thing that um if I had to boil down the biggest change in risk and why it matters now more than ever, it's outputs to actions, right?
00:14:04.879 --> 00:14:09.679
What we had to worry about a couple of years ago from AI in terms of risk was the outputs.
00:14:09.840 --> 00:14:16.799
Now we have to worry about the actions, but it's actions at scale and actions that we might not even necessarily be able to see, right?
00:14:17.039 --> 00:14:21.840
Silent, unintended workflows that you may not be tracing tracing.
00:14:22.879 --> 00:14:33.360
And really what this comes down to, well, it's this combination of increased capabilities from agents and moving in the shadows, and that's an enterprise nightmare.
00:14:33.440 --> 00:14:36.080
That is literally the formula for an enterprise nightmare.
00:14:36.240 --> 00:14:38.000
So, some stats here for you, right?
00:14:38.080 --> 00:14:44.159
So, right now, 57% of employees uh at least admit to using personal AI accounts for work.
00:14:44.240 --> 00:14:45.759
It's way more than that, let's be honest.
00:14:45.840 --> 00:14:47.200
That's just the number that admit.
00:14:47.279 --> 00:14:51.919
Uh, a third admit to inputting sensitive data into unapproved tools, right?
00:14:52.000 --> 00:14:54.159
So your shadow AI use case there.
00:14:54.320 --> 00:14:59.840
And here's the thing you can't govern what you can't see.
00:15:00.720 --> 00:15:06.480
And most organizations can't see their agent footprint at scale.
00:15:08.639 --> 00:15:12.320
This is what I call the three types of dark AI.
00:15:12.480 --> 00:15:12.799
All right.
00:15:12.879 --> 00:15:22.480
For the most part, you're not gonna find you know the three dark types of or the three types of dark AI online is something that I've kind of uh categorized them in.
00:15:22.559 --> 00:15:28.080
But I think it's really helpful uh to get a better glimpse and a better understanding of the categories of risk.
00:15:28.240 --> 00:15:31.279
So, number one, we all know this, shadow AI, right?
00:15:31.360 --> 00:15:35.679
That's just essentially unapproved or unknown AI use, right?
00:15:35.919 --> 00:15:45.600
If you've been using, you know, Chat GPT on your personal computer uh because Copilot is the AI that's approved, but you want to use Chad GPT and you copy and paste things over as an example, right?
00:15:45.679 --> 00:15:47.679
That's shadow AI, but that's been around.
00:15:47.759 --> 00:15:49.039
Everyone knows that, right?
00:15:49.200 --> 00:15:54.480
Uh what you've maybe heard of, maybe not, it's the next kind of tier, and that's agent sprawl.
00:15:54.720 --> 00:15:58.080
But agent sprawl is known, right?
00:15:58.240 --> 00:16:06.159
So that's essentially when you have approved agents, but you're not sure how to wrangle them or observe them all.
00:16:06.240 --> 00:16:15.440
You're like, oh, well, yeah, Bill gave us that, you know, agent to help with finances, but we're not really sure what it's doing, right?
00:16:15.919 --> 00:16:18.000
Um, we think it's doing good, right?
00:16:18.080 --> 00:16:20.879
I checked the outputs, but I don't really know how it's getting there.
00:16:21.120 --> 00:16:23.840
That's that's that's the beginning of agent sprawl.
00:16:24.000 --> 00:16:27.759
But the thing is, agent sprawl goes quickly.
00:16:28.000 --> 00:16:34.639
In the same way, a snowball at the top of the mountain might come rumbling down a thousand times the size.
00:16:34.799 --> 00:16:38.879
That is where we get into then dark agent sprawl.
00:16:39.120 --> 00:16:39.759
All right.
00:16:40.080 --> 00:16:44.879
This sounds like a like a screen name for uh for aim back in the back in the 90s.
00:16:45.120 --> 00:16:46.080
You guys remember that?
00:16:46.240 --> 00:16:46.559
Right?
00:16:46.720 --> 00:16:49.679
AIM, uh, you know, instant messenger, right?
00:16:49.919 --> 00:16:53.919
Dark Agent Sprawl, uh 2026.
00:16:54.000 --> 00:16:59.519
Uh that's gonna be my uh my username on the uh everyday AI inner circle community.
00:16:59.679 --> 00:17:05.359
Um but essentially that's where there's well dark agent sprawl can be a couple of things because agent sprawl, it's a problem.
00:17:05.440 --> 00:17:07.839
But like I said, for the most part, that's a known risk.
00:17:08.000 --> 00:17:11.759
And you're like, yeah, we have these agents going all over the place.
00:17:12.240 --> 00:17:18.480
We we have no guardrails, we have no uh traceability, observability, right?
00:17:18.720 --> 00:17:20.559
It's but but you know of it.
00:17:20.640 --> 00:17:30.400
Dark agent sprawl is agents you don't know about, those are unapproved agents that are working in or on your company, and you can't observe them because you don't know about them.
00:17:30.480 --> 00:17:35.680
So this is well, this can be shadow AI, you know, uh gone agency, right?
00:17:35.759 --> 00:17:41.839
So people plugging in agents that aren't approved and company doesn't know, but there's also another kind, right?
00:17:42.160 --> 00:17:46.880
Uh and I think we're gonna see a lot of this, maybe not in 2026, but in 2027.
00:17:46.960 --> 00:17:52.559
That's going to be the equivalent of malware or spyware, but agents, right?
00:17:52.640 --> 00:18:07.440
People sending uh and seeding agents out specifically in the same way that they would for spyware, uh, malware, ransomware to make money, uh, right, to extract uh, you know, value from businesses.
00:18:07.519 --> 00:18:08.960
And that's what's gonna happen.
00:18:09.119 --> 00:18:11.200
Um, that's the next version of this.
00:18:11.279 --> 00:18:14.319
That's not necessarily dark agent sprawl, but there's two sides, right?
00:18:14.480 --> 00:18:17.359
Dark agent sprawl can start out innocent enough, uh, right?
00:18:17.519 --> 00:18:21.519
Oh my gosh, you know, our company's not approving, you know, any of my AI agents.
00:18:21.599 --> 00:18:22.559
Well, okay, whatever.
00:18:22.720 --> 00:18:32.160
I'm gonna go ahead and um, you know, unleash uh, you know, clawed code and get, you know, 50 instances of clawed code going up and they're all gonna spin their agents.
00:18:32.240 --> 00:18:34.960
Well, one person might know, but the rest of the company is in the dark.
00:18:35.119 --> 00:18:38.480
And those agents could replicate, duplicate, and you can't observe them.
00:18:38.559 --> 00:18:43.680
But the other thing is, well, bad actors, dark agent sprawl, that's a thing too.
00:18:44.079 --> 00:18:46.480
So why is this all happening now?
00:18:46.720 --> 00:18:50.240
All these risks and security concerns and the sprawl, why why now?
00:18:50.400 --> 00:18:54.400
Because, like I've said, we've been six months away for five years, right?
00:18:54.480 --> 00:18:59.440
We're six months away from great agents, but then it's all it almost seems like we didn't get a six-month warning.
00:18:59.680 --> 00:19:01.440
They just popped up, right?
00:19:01.759 --> 00:19:03.759
Like December 2025, right?
00:19:03.839 --> 00:19:08.799
Like everyone was winding down for you know the holidays, you know, especially here in the US.
00:19:08.880 --> 00:19:12.720
And then you go back to work in January, you're like, what the frick happened, right?
00:19:13.200 --> 00:19:17.279
You're like, it's like, whoa, whoa, wait, the the agents are actually here now.
00:19:17.359 --> 00:19:18.559
Where's the six-month warning?
00:19:18.720 --> 00:19:21.519
Why don't we get get word of this in in June or July?
00:19:21.680 --> 00:19:23.359
But there's three reasons why.
00:19:23.440 --> 00:19:31.599
And it is literally the perfect storm, not just the ingredients happening, but at the exact right or wrong time.
00:19:31.759 --> 00:19:32.000
All right.
00:19:32.079 --> 00:19:40.000
So, number one, the number one element uh that was mixed in the bowl and it's exploding is the reasoning threshold.
00:19:40.400 --> 00:19:40.880
Right.
00:19:41.039 --> 00:19:58.799
So improved reasonings from models like you know, GPT-5.3.1, uh, their new model, very impressive, and you know, Opus Sonnet 4.6 from Anthropic, these are all built to be agent native.
00:19:59.279 --> 00:20:01.519
That's how they build them now, right?
00:20:01.759 --> 00:20:06.480
They're not building models that are great at you know, reading, writing, and comprehension first.
00:20:06.880 --> 00:20:11.680
No, what they're focusing on is is the harness and and the tool use.
00:20:11.759 --> 00:20:12.799
That's what's first, right?
00:20:12.880 --> 00:20:19.759
Even you know, Google's um, you know, their new model yesterday, everything that they highlighted was tool use, right?
00:20:19.920 --> 00:20:27.359
And and and how, you know, improved tool use and the scaffolding there has, you know, really helped them um improve their outputs.
00:20:27.839 --> 00:20:29.119
But that's that's the thing.
00:20:29.279 --> 00:20:32.480
These models, their reasoning ability is legit through the roof.
00:20:32.559 --> 00:20:39.519
They can plan steps ahead, uh, self-correct errors, and move beyond uh reactive behavior into proactive.
00:20:39.599 --> 00:20:43.839
And that's kind of reduced agent reliability from around 50% to 90%.
00:20:44.240 --> 00:20:46.880
And that's perfect storm right there, right?
00:20:47.039 --> 00:20:49.599
50% coin flip, you're not doing that, right?
00:20:49.839 --> 00:20:52.880
Would you hire an employee that has a 50% chance to fail right away?
00:20:53.039 --> 00:20:54.960
Probably not, but 90%, okay.
00:20:55.119 --> 00:20:56.960
That's an A employee.
00:20:57.519 --> 00:21:00.640
Um, step two, computer use improvements.
00:21:00.720 --> 00:21:02.880
This seems small, but this is big.
00:21:03.039 --> 00:21:03.359
Okay.
00:21:03.759 --> 00:21:06.559
The models have to be insanely smart, right?
00:21:06.640 --> 00:21:08.079
They have to be able to think, reason.
00:21:08.160 --> 00:21:11.440
They have to have genius level IQ, which is what today's models do, right?
00:21:11.519 --> 00:21:17.519
If you look at offline IQ tests, they're scoring in the genius level, smarter than 99.9% of people, right?
00:21:17.599 --> 00:21:19.200
Uh, but the computer use is important.
00:21:19.279 --> 00:21:25.759
It doesn't matter if it can't go and use a computer, but because at least for now, that's how businesses generate value.
00:21:25.839 --> 00:21:29.920
I think eventually value will be just generated agentically, right?
00:21:30.079 --> 00:21:32.319
The web has gone agentic, right?
00:21:32.480 --> 00:21:39.920
Google and Microsoft uh announced support for essentially an MCP version of the web, um, where websites can talk to each other, agents can talk to each other.
00:21:40.079 --> 00:21:43.279
So we'll see how you know business value is extracted in the future.
00:21:43.519 --> 00:21:49.119
Because a lot of times it's been through the website stack, the software stack, but you know, who knows what it'll be in the future?
00:21:49.200 --> 00:21:53.200
But right now, agents can use computers better than humans, right?
00:21:53.359 --> 00:21:59.839
So uh computer use means uh these models can use a mouse computer, they can uh click, they can use API.
00:22:00.480 --> 00:22:01.680
They can talk to each other.
00:22:02.000 --> 00:22:05.920
But the big thing is the computer use capability gap has essentially been solved.
00:22:06.240 --> 00:22:10.319
So Claude Sonnet 4.6, which came out well earlier this week.
00:22:10.480 --> 00:22:10.799
All right.
00:22:10.960 --> 00:22:14.960
So if you're listening to this in six months, right, it came out in uh mid-February.
00:22:15.039 --> 00:22:19.920
So imagine in six months, uh, these these uh computer use models are gonna be uh insanely good.
00:22:20.000 --> 00:22:26.079
Uh but it's the first um the first one that scored better than humans, right?
00:22:26.160 --> 00:22:37.759
So 72.5 uh success rate on the OS World benchmark, and that's surpassing human performance for the first time and nearly quintupling the scores of 2024.
00:22:37.839 --> 00:22:38.240
That's the thing.
00:22:38.319 --> 00:22:48.559
And I think that's one of the reasons why, you know, even though the uh maybe the foundation was there for AI agents uh, you know, a year, year and a half ago, they couldn't use the computer.
00:22:48.640 --> 00:22:50.480
It was so slow, uh, right?
00:22:50.640 --> 00:22:53.759
It was, you know, very, very, very slow computer vision, right?
00:22:53.839 --> 00:22:57.279
Every time, if you wanted to click something, it would take three minutes.
00:22:57.440 --> 00:22:57.759
Okay.
00:22:58.079 --> 00:23:08.960
Now models can go as fast as humans, or at least, you know, uh not just clicking and navigating interfaces, but just uh a range of computer use tasks.
00:23:09.599 --> 00:23:15.839
And then the third thing that has created this perfect agentic storm, well, it's the context window in memory, uh, right.
00:23:15.920 --> 00:23:20.319
So being able to work on tasks for a long period of time, uh right.
00:23:20.480 --> 00:23:30.319
I think my uh my record uh so far, I think I got to about 10 hours uh overnight the other night on codex, um, which was really fun to do, right?
00:23:30.559 --> 00:23:34.160
But it kept its memory persistent the whole time, right?
00:23:34.319 --> 00:23:36.559
It didn't forget what I told it when I went to wake up.
00:23:36.720 --> 00:23:39.839
You know, when I woke up, yeah, I spent time reading through the chain of thought.
00:23:39.920 --> 00:23:40.799
And I'm like, oh, sweet.
00:23:40.880 --> 00:23:48.720
Not only did it remember anything, but I'm going through, I'm, you know, uh tracing it, I'm, you know, making sure it kind of stayed within the confines of the instructions I gave it.
00:23:48.799 --> 00:23:49.279
And it did.
00:23:49.359 --> 00:23:51.039
But, you know, that's another reason why it's happening.
00:23:51.119 --> 00:24:13.519
And y'all, let me tell you, if you haven't in the last like three, four weeks, if you haven't gone out and used ChatGPT or you know, OpenAI's codex, if you haven't used um, you know, Claude Code or Claude Cowork from Anthropic, if you haven't used uh Anti-Gravity uh from Google, I'm not saying this to like be the guy, you you're gonna get left behind, right?
00:24:13.680 --> 00:24:16.240
And if you're listening to this show, I don't want you left behind.
00:24:16.319 --> 00:24:19.039
I'm actually thinking, all right, I don't know this yet.
00:24:19.200 --> 00:24:24.640
It might might kind of um get a little bit of a vibe working focus.
00:24:24.880 --> 00:24:27.759
Um, you you know, here in the uh first or second quarter.
00:24:27.839 --> 00:24:32.319
So uh make sure if you're not already in our uh inner circle community, get in there.
00:24:32.480 --> 00:24:35.839
Um yeah, I'm gonna start putting putting some stuff in there.
00:24:35.920 --> 00:24:38.160
Just keep keep uh keep your eyes open for that.
00:24:38.480 --> 00:24:43.279
All right, but let's keep going a little bit here as we uh get gonna be wrapping up here in a minute.
00:24:43.359 --> 00:24:52.400
Uh, let's talk a little bit about how the biggest AI companies are actually responding right now because the risk is well known because it has literally been an explosion and it is going to get worse.
00:24:52.559 --> 00:24:57.200
So I want to talk about a little bit about kind of the uh big picture focus of the big four labs.
00:24:57.279 --> 00:25:01.279
So right now, OpenAI is taking more of a human approval uh approach, right?
00:25:01.440 --> 00:25:05.119
So Codecs is kind of their command center uh to review agent decisions.
00:25:05.359 --> 00:25:22.400
Anthropic is really going on the defense against prompt injection uh for browser agents, um, you know, with uh making sure uh that virtual machines are isolated and have minimal privileges, uh, and going with domain allowed list versus blacklists, right?
00:25:22.480 --> 00:25:24.880
So everyone has a little bit of a different uh approach here.
00:25:24.960 --> 00:25:34.559
You know, Google, as an example, there, uh Project Mariner, uh browser agents run in virtual machines as a safety uh to isolate them from anything else they can kind of get their hands on.
00:25:34.640 --> 00:25:35.839
And then Microsoft, right?
00:25:35.920 --> 00:25:39.839
There's millions of ways that Microsoft is doing it and millions of ways that all these companies are doing it.
00:25:39.920 --> 00:25:44.000
I just kind of picked out uh you know different aspects to illustrate their different approach.
00:25:44.160 --> 00:25:49.519
You know, Microsoft, as an example with their co-pilot studio, I mean the governance is everywhere, right?
00:25:49.599 --> 00:25:56.000
They have the Sentinel monitoring, the uh the purview logging, you know, the agent uh intra ID, uh, right.
00:25:56.079 --> 00:25:59.440
So Microsoft, it is very enterprise, Google is as well.
00:25:59.519 --> 00:26:04.640
I was just giving an example of you know how their project mayor kind of runs in an isolated virtual machine.
00:26:04.799 --> 00:26:18.319
Uh because essentially, I think from the people I've talked to, um, at least uh three of those four companies, um they know risk is part of it, right?
00:26:18.640 --> 00:26:21.920
Like there's like any lab, right?
00:26:22.079 --> 00:26:25.599
They call them frontier, you know, or AI labs for a reason.
00:26:25.759 --> 00:26:31.119
You run experiments, and part of it is well, you know that things need to go wrong, right?
00:26:31.200 --> 00:26:39.200
So these companies are intentionally trying to get all the risk and all the security nightmares and all the sprawl they can, right?
00:26:39.279 --> 00:26:45.119
So then when they set a new model out in the wild, when they're done with the red teaming, um, you know, they hope that they have a good idea.
00:26:45.279 --> 00:27:02.640
So the labs are working on this, but here's the reality I think there's a pressure to allocate more resources to the development of models versus research and security, right?
00:27:03.359 --> 00:27:06.319
I don't have that on you know, privileged and information.
00:27:06.400 --> 00:27:12.160
That's just from you know, talking to a lot of people and reading a lot and just the I think the reality of the world that we live in.
00:27:12.799 --> 00:27:19.279
And that's why AI sprawl is going to be hard to miss for most business leaders.
00:27:19.440 --> 00:27:22.480
Um, because everything's unclear, right?
00:27:22.559 --> 00:27:24.079
The tool, like, do you know?
00:27:24.240 --> 00:27:26.880
Think of the the the AI model you use most.
00:27:27.599 --> 00:27:28.880
Do you know the tools it has?
00:27:29.519 --> 00:27:29.759
Right.
00:27:29.920 --> 00:27:36.640
Even if you think of, you know, Chat GPT, Gemini, Claude, uh, do you know what tools the base models have?
00:27:37.279 --> 00:27:38.240
Probably not.
00:27:38.960 --> 00:27:40.559
You'd be surprised, right?
00:27:40.720 --> 00:27:46.160
Like, I'm one of those weird guys that that reads uh change logs um and model cards.
00:27:46.319 --> 00:27:48.400
Most people have no clue, right?
00:27:48.640 --> 00:28:00.960
And this is much worse than classic shadow IT because agents can act across systems, not just you know, within the confines of uh folders and files, which are very structured.
00:28:01.200 --> 00:28:08.160
The whole kind of um, it's kind of like how people talk about hallucinations, right?
00:28:08.319 --> 00:28:11.200
Um, oh, you know, they're not a bug, they're a feature, right?
00:28:11.359 --> 00:28:15.680
What makes agents great is also the built-in nightmare.
00:28:15.759 --> 00:28:20.960
So if you want to have the good, you have to recognize the bad that they can do.
00:28:21.200 --> 00:28:22.000
Um, right.
00:28:22.079 --> 00:28:23.759
And that's because that's how they're made.
00:28:23.920 --> 00:28:29.200
They are made to go build their own path, they are made to blaze their own trail, right?
00:28:29.359 --> 00:28:35.839
So by their very definition, agents aren't necessarily always good at staying within guardrails, right?
00:28:35.920 --> 00:28:44.480
Because if they think that the destination that has been given to them, um, at times, right, depends on the model to set up, all that, right?
00:28:44.640 --> 00:28:53.599
But it's very common for an agent to hop over guardrails in order to accomplish a task, not knowing that that's a bad thing, right?
00:28:54.000 --> 00:29:05.839
And every new skill that you add that you give to uh an agent or a model, every connector, every workflow, uh, all that does is it expands the attack surface for everyone.
00:29:06.240 --> 00:29:12.079
And that's another realization that I think most people are going to struggle to grasp, right?
00:29:12.160 --> 00:29:18.240
Because you think that you're expanding the amount of work that you can get done.
00:29:18.480 --> 00:29:19.440
And are you?
00:29:19.759 --> 00:29:20.240
Sure.
00:29:20.400 --> 00:29:26.319
But I think you almost have to think of it from like a, I don't know, like an old school like military perspective, right?
00:29:26.400 --> 00:29:27.680
It's your land, right?
00:29:27.839 --> 00:29:29.680
And hey, we're gonna go discover new land.
00:29:29.759 --> 00:29:32.880
We're gonna go out and, you know, win new deals, right?
00:29:33.039 --> 00:29:34.559
So we're gonna expand our land.
00:29:34.640 --> 00:29:37.279
I don't know if any of you have ever played like, you know, Risk.
00:29:37.759 --> 00:29:38.960
Um, I used to.
00:29:39.119 --> 00:29:40.319
I wish I still had time, right?
00:29:40.400 --> 00:29:44.799
It's one of those games, you know, it's like, oh, okay, this is good to make sure my brain still works.
00:29:44.880 --> 00:29:46.400
But it's kind of like that, right?
00:29:46.559 --> 00:29:51.279
Um, your agents, without you knowing, though, are acquiring new territories.
00:29:51.440 --> 00:29:52.240
And that's great.
00:29:52.319 --> 00:29:55.680
And you think, oh my gosh, look at these new skills and capabilities, right?
00:29:55.759 --> 00:29:58.400
It's like, oh, it's like having a thousand new employees.
00:29:58.559 --> 00:29:59.200
Okay, great.
00:29:59.359 --> 00:30:08.799
But your surface area for risk, attacks, and sprawl is also multiplying at the same time as your capability.
00:30:08.960 --> 00:30:10.319
But you're not focused on that.
00:30:10.480 --> 00:30:13.759
You're focused on, oh my gosh, now all of a sudden I have a designer.
00:30:13.839 --> 00:30:15.839
Now all of a sudden I have a data analyst.
00:30:16.000 --> 00:30:20.079
Okay, well, what's that data analyst doing with your data?
00:30:20.400 --> 00:30:21.920
Are you going to check?
00:30:22.240 --> 00:30:22.559
Right?
00:30:22.799 --> 00:30:25.440
Is it uh using an in open source, right?
00:30:25.519 --> 00:30:28.319
This is one thing I was kind of blown away by the other day.
00:30:28.720 --> 00:30:33.599
Uh gave, gave uh, you know, it wasn't anything bad, right?
00:30:34.000 --> 00:30:35.839
Gave a model a very hard task.
00:30:36.000 --> 00:30:45.519
This is one of those that uh overnight um I came back and I realized, oh, it downloaded another large language model locally to do the task, right?
00:30:45.599 --> 00:30:47.359
It did it locally, uh, right.
00:30:47.440 --> 00:30:52.319
So uh kept all the data uh kind of on on-prem on my local machine.
00:30:52.640 --> 00:30:59.200
But what happens if in the future other AI models are just gonna go use other AI models, but on the web, right?
00:30:59.359 --> 00:31:00.079
They can.
00:31:00.319 --> 00:31:02.319
What if they do that without you telling?
00:31:02.480 --> 00:31:06.240
Or what if they go use a um a site that's unsecure?
00:31:06.319 --> 00:31:13.200
And and that, and this is why I think, you know, open claw, as amazing as it is, this is why it's also scary, right?
00:31:13.599 --> 00:31:26.000
Because, you know, these you know, uh autonomous open source agents, the open source, if I'm being honest, and I'm sure, you know, uh I'm gonna catch some flack from this from some OG open source people, right?
00:31:26.160 --> 00:31:29.200
Um, OG open source, I'm not talking about you.
00:31:29.359 --> 00:31:35.039
Today's open source, it's taken a weird turn over the past uh three to six months, right?
00:31:35.200 --> 00:31:42.160
It's almost become this, you know, crypto-infused, wild, wild web point four west, right?
00:31:42.240 --> 00:31:45.279
It's not like traditional open source anymore.
00:31:45.359 --> 00:31:53.119
So when we talk about open source AI agents, right, so many of what's so much of what's out there is risk, right?
00:31:53.359 --> 00:31:58.559
It is, and it's unknown and it can be scary if you don't know exactly what you're getting into.
00:31:59.680 --> 00:32:04.400
And I think this is why small agent pilots are becoming huge risk, right?
00:32:04.559 --> 00:32:08.880
Because when you build your own agents without centralized visibility or governance, that's where you get in trouble.
00:32:08.960 --> 00:32:12.319
So many people and they're using their own agents, it's for them, right?
00:32:12.559 --> 00:32:18.960
But it can be go going and doing work for the whole department, the whole company, or accessing data from the whole company.
00:32:19.039 --> 00:32:26.640
But if one person has eyes on it, and if there's not central um organization, that's where you run into something.
00:32:27.519 --> 00:32:29.759
So here's the playbook, right?
00:32:30.079 --> 00:32:33.039
If you're feeling a little scared, you know what?
00:32:33.119 --> 00:32:34.319
Maybe that was my intent.
00:32:34.480 --> 00:32:44.000
I got y'all pumped up and um, you know, excited with yesterday's show talking about, you know, the new capabilities of agents, but you gotta get it in check, right?
00:32:44.160 --> 00:32:47.279
You have to balance that optimism with a little bit of realism.
00:32:47.359 --> 00:32:52.960
But here is the Monday morning playbook for how to deal with the risk, the attacks, and the sprawl.
00:32:53.039 --> 00:32:58.400
Uh so like what we talked about yesterday, touched on it briefly, but you need to start with bounded autonomy, right?
00:32:58.640 --> 00:33:04.000
When you think of autonomous agents, it's not uh, you know, zero to a hundred.
00:33:04.079 --> 00:33:05.200
It's bounded autonomy.
00:33:05.279 --> 00:33:13.920
What that means is you start with suggesting, then propose, then approve, then limited um execution, right?
00:33:14.079 --> 00:33:16.480
So many people go to full execution first.
00:33:16.960 --> 00:33:18.720
It's baby steps, right?
00:33:19.039 --> 00:33:23.359
Uh a human being doesn't go from womb to sprint, right?
00:33:23.440 --> 00:33:33.359
You go from womb to you know being being on your stomach to sitting, to crawling, to walking, to right?
00:33:33.759 --> 00:33:35.359
Your agents have to go the same way.
00:33:35.440 --> 00:33:41.759
Even if you know out of the box that agent can sprint, you're not prepared for that agent to sprint.
00:33:41.920 --> 00:33:43.759
Part of it is for the humans, right?
00:33:43.920 --> 00:33:47.920
We think it's more for um understanding the the agent's abilities.
00:33:48.000 --> 00:33:49.680
It's not, right?
00:33:50.079 --> 00:33:53.200
This is to make sure that you don't have a bunch of lazy humans in the loop.
00:33:53.279 --> 00:33:57.599
And instead, you have uh proactive expert-driven loops.
00:33:57.759 --> 00:34:00.720
Um so start with least privilege by default.
00:34:00.880 --> 00:34:05.599
Read only first, write access only for narrow defined tasks, right?
00:34:05.759 --> 00:34:10.800
Especially in enterprise environments, you don't want to send out a bunch of read-only general agents.
00:34:11.039 --> 00:34:14.960
Start with read only first, observe, uh, improve your loop.
00:34:15.039 --> 00:34:21.440
Then you go to um limiting, uh, limited execution.
00:34:21.760 --> 00:34:28.159
Um, and then you can finally get it to um writing for narrow tasks.
00:34:28.320 --> 00:34:38.159
Uh, and then last in your Monday morning playbook is you need to require human approvals for irreversible actions like sends, deletes, purchases, and permission changes.
00:34:38.400 --> 00:34:40.800
These are all things agents can do, right?
00:34:40.960 --> 00:34:42.559
Agents can literally do anything.
00:34:42.719 --> 00:34:52.559
Agents, uh, right, and this is going to get even harder to manage when we start to see true agentic commerce, right?
00:34:52.639 --> 00:34:59.519
And I'm not just saying, oh, you know, my agent is gonna go buy something, you know, from uh the Amazon agent.
00:34:59.679 --> 00:35:01.039
That's not what I'm saying.
00:35:01.119 --> 00:35:01.599
Uh, right.
00:35:01.679 --> 00:35:08.960
There's going to be agent uh bartering, there's going to be agent um, you know, liaisons, right?
00:35:09.039 --> 00:35:13.760
So uh you have to work on human approvals now, right?
00:35:13.920 --> 00:35:29.280
It's almost like you have to understand that agents are capable of you know a 10 and you have to give them a two because we as humans need to first learn to adapt our behavior um before just giving in to agents, right?
00:35:30.000 --> 00:35:33.360
And then you need to build governance before you scale.
00:35:33.599 --> 00:35:36.079
And I think that's like just what I was getting at.
00:35:36.239 --> 00:35:36.480
Right.
00:35:36.719 --> 00:35:40.800
We get all excited for what we know it can do.
00:35:40.880 --> 00:35:43.360
And then we're like, wait, we got to do this boring stuff first?
00:35:43.599 --> 00:35:44.400
Absolutely.
00:35:44.639 --> 00:35:52.000
Because if you don't do the boring stuff first, if you don't go through the sitting on your tummy, right?
00:35:52.800 --> 00:36:00.800
Sitting up, crawling, walking, then you can't expect to understand uh the path of the sprinting agent.
00:36:00.960 --> 00:36:04.719
And every agent run needs a decision trace that you can expect after the fact.
00:36:04.800 --> 00:36:11.039
So you have to be able to log all your tool calls, capture your decision traces, and monitor for abnormal action patterns.
00:36:11.199 --> 00:36:18.719
This is one of the reasons why I said, I think we're gonna see in 2026, uh, it's gonna be extremely common to have agent op teams, right?
00:36:18.800 --> 00:36:23.280
In the same reason, or sorry, in the same vein, you know, dev op teams are very common, right?
00:36:23.360 --> 00:36:26.480
We're gonna have agent op teams for this exact reason.
00:36:26.960 --> 00:36:32.880
So here's what to watch for for the rest of 2026 uh when it comes to agent risk security and sprawl.
00:36:33.119 --> 00:36:36.400
Uh browser agents are gonna become a mainstream risk surface.
00:36:36.480 --> 00:36:37.360
I think we know that.
00:36:37.440 --> 00:36:40.400
Uh, I think we're gonna see a major open claw type.
00:36:40.480 --> 00:36:45.519
I'm not saying open claw themselves, but I think we're gonna see a major uh open source agent crash.
00:36:45.679 --> 00:36:48.559
I talked about that in the 2026 AI prediction and roadmap series.
00:36:48.719 --> 00:36:50.320
I actually think it's gonna come from, right?
00:36:50.400 --> 00:36:59.440
There's this new trend, like trend over the last like week or two, where everyone is just sending pointing their agents to a uh a website, right?
00:36:59.599 --> 00:37:12.159
When they want their agent to do something, they're not even saying, all right, let me uh let me, the human, understand this, let me write it out, let me apply this to my business logic, let me apply this with my garden.
00:37:12.239 --> 00:37:12.480
Nope.
00:37:12.559 --> 00:37:17.039
They're just saying pointing their agent, hey agent, here's this video, go watch it and do it.
00:37:17.119 --> 00:37:22.079
Hey, agent, here's this website with a whole bunch of cool stuff, just go do it.
00:37:22.320 --> 00:37:26.880
You know, there's this whole concept now going around of, you know, point your agent to a URL.
00:37:27.599 --> 00:37:29.280
Okay, have fun with that.
00:37:29.679 --> 00:37:31.360
Don't do that, right?
00:37:31.599 --> 00:37:32.400
Don't, right?
00:37:32.480 --> 00:37:35.440
Especially if that agent has its own dedicated machine, right?
00:37:35.519 --> 00:37:38.320
And I'm saying if you're tinkering around, right?
00:37:38.400 --> 00:37:41.679
It's it's it's your personal computer, you're the business owner.
00:37:41.760 --> 00:37:43.360
If you want to take that risk, that's fine, right?
00:37:43.440 --> 00:37:45.840
Especially if you're you're you're tinkering around, that's fine.
00:37:45.920 --> 00:37:47.679
But don't do that in an enterprise.
00:37:47.840 --> 00:37:52.960
And uh unfortunately, people are doing that, and that's a bad thing because guess what?
00:37:53.119 --> 00:37:57.599
A lot of those sites, quote unquote, oh, you know, someone puts up a huge tutorial.
00:37:57.679 --> 00:37:59.119
Uh okay.
00:38:00.079 --> 00:38:01.920
They're putting it up on their personal website.
00:38:02.000 --> 00:38:06.239
You know how easy it is to hack, uh, to fish any of these websites?
00:38:06.480 --> 00:38:08.239
Okay, you see, let me give you an example.
00:38:08.400 --> 00:38:09.679
Something goes viral on Twitter.
00:38:09.840 --> 00:38:16.719
It's some, you know, uh literally, some 20-year-old kid, you know, put a great uh, you know, guide up on the website.
00:38:16.880 --> 00:38:17.119
Cool.
00:38:17.199 --> 00:38:20.559
And, you know, everyone's just like, hey, agent, go read that, go read that.
00:38:20.639 --> 00:38:25.039
And then you have millions of agents going to just read this, you know, this nice kid, put up a cool guide.
00:38:25.280 --> 00:38:26.960
Okay, guess what's gonna happen?
00:38:27.119 --> 00:38:32.639
Someone's gonna see that, they're gonna hack that site and they're gonna inject malicious code in there that no one's gonna see.
00:38:32.719 --> 00:38:33.679
That's what's gonna happen.
00:38:33.920 --> 00:38:38.719
All right, next, the agent skill marketplace is going to expand uh the supply chain risk.
00:38:38.800 --> 00:38:40.400
So you inherit what you plug in.
00:38:40.480 --> 00:38:44.400
And we've already seen that with uh some skills and plugins for agents.
00:38:44.480 --> 00:38:48.559
Uh, some of the top ones were found to have uh you know malicious code in them.
00:38:48.639 --> 00:38:55.119
Uh, you need to uh we're gonna see identity and permissions becoming a board level compliance requirement, not a nice to have.
00:38:55.280 --> 00:39:01.039
And last but not least, the winners are gonna treat agents like production software, not side experiments.
00:39:01.199 --> 00:39:02.159
All right, that's it.
00:39:02.239 --> 00:39:05.360
That's a wrap volume nine of the start here series.
00:39:05.599 --> 00:39:06.480
I hope this is helpful.
00:39:06.559 --> 00:39:27.679
I hope this was helpful and I hope you understand a little bit more the risk security and the sprawl uh that you need to be aware of and how this having AI that acts, having AI, right, that has a smart brain, it's it's proactive, it has tools, and it has arms.
00:39:28.559 --> 00:39:30.480
That changes everything, right?
00:39:30.559 --> 00:39:42.559
This perfect storm that has happened in the last 30 days, we've seen, I've said this before, we've seen more both on the opportunity and the risk side.
00:39:42.719 --> 00:39:47.920
We've seen more in the last 30 days than we've seen in the last probably two years.
00:39:48.159 --> 00:39:48.719
All right.
00:39:48.960 --> 00:39:54.320
So I wanted to take a moment for the start here series to address this.
00:39:54.480 --> 00:40:03.920
And I hope this helps you not only understand the risk, but if you want to take advantage of the opportunities, you can't do it without knowing the risk.
00:40:04.079 --> 00:40:04.719
So now you do.
00:40:04.880 --> 00:40:06.480
All right, so I hope this was helpful.
00:40:06.639 --> 00:40:12.719
So if so, there's already eight other uh in our start here series that you should go check out.
00:40:12.800 --> 00:40:15.920
So, like I said, please go to starthearseries.com.
00:40:16.000 --> 00:40:19.360
That is gonna give you free access to our inner circle community.
00:40:19.519 --> 00:40:34.079
Once you sign up, you're gonna be uh just uh straight up loving it, hopefully, uh, in our start here series space and connecting uh with now more than a thousand people uh in our uh inner circle community.
00:40:34.239 --> 00:40:35.360
So thank you for tuning in.
00:40:35.519 --> 00:40:38.320
Hope to see you back for more everyday AI.
00:40:38.400 --> 00:40:39.360
Thanks, y'all.
