Episode Transcript
[00:00:00] Sa, Here's a true sentence about the software that you're running in your business.
[00:00:36] It can move money, can sign an agreement, can push code to production, publish anything to the public, and even open an account all on its own, even maybe while you sleep.
[00:00:51] And here's the question, though, that almost nobody built a way to answer before they turned it on.
[00:01:01] Who authorized that?
[00:01:04] I'm Dr. Alan Badot, this is AI Today. And you know, this week we're going to deep dive into a discussion like this. You know, last week we, we went underneath the model, if you remember, right, how the power feeds it and permission, you know, would let it run and those kind of things. And, you know, but this week we're going in the other direction. We're going to go up, we're going to look at the agents that are now standing on, on top of the model, taking actions that have real implications in the real world.
[00:01:39] And you know, the oldest idea in any serious organization now suddenly the most urgent idea in technology is, is pushing that. And it's the chain of command.
[00:01:56] So every action in a functioning institution can be traced, right? A soldier acts on an order. An order comes from an officer. The officer orders, you know, and answers to a commander, and the commander to, you know, somebody else above him and the, you know, somebody else above them, a general maybe, to, you know, the law, right?
[00:02:20] You can walk that chain link by link from the action all the way back to the human who is, you know, accountable for it. And, you know, that chain's not bureaucracy, okay? That chain is how a powerful thing stays legitimate.
[00:02:43] So, you know, this week the AI industry discovered all at once that it's been handing enormous power to agents while leaving that chain half built.
[00:02:57] The agent acts.
[00:02:58] And when you go walk that chain back, you look at who authorized this and by whose order or, you know, who answers if it's wrong, and you find that the links just not all there.
[00:03:14] Now, tonight, I'm going to do this, you know, as honestly as I possibly can, right? Because we talk about, I'm doing the exact same thing. We're building these things, and I'm going to put my perspective on the things, but I'm going to give you both sides, okay? Because I don't want you to think that this is a sales pitch for agents, you know, and it's also not a panic about them either, okay? So we're going to do the good, the bad and maybe the ugly, you know, all three and take a look at them because all three are real.
[00:03:49] And you deserve to See the whole picture before you wire one of these things potentially into your, into your bank account, because the good is genuinely good.
[00:04:01] You know, the money tells you quite honestly how good that that really is. The analyst at Gartner now put spending on agent software at more than $200 billion this year.
[00:04:17] That's right, $200 billion. And that's a jump of 139% over last year, which makes it, of course, the fastest growing slice of enterprise software there is.
[00:04:32] And this is not hype money, okay? This is, this is budget money signed off by people who expect a return on that.
[00:04:43] Let me say that one more time for you.
[00:04:46] $206 billion this year on agent related software spend, and that's 139% increase over last year.
[00:04:58] That's impressive, right?
[00:05:00] And you know, the capability under that number, it's, it's real. You know, this week, if you look at, you know, a platform from the BNB chain and Amazon, they, you know, of course, let a developer stand up an autonomous agent, and it had its own wallet, its own ability to transact things, and from a single prompt in about 15 minutes.
[00:05:26] Yeah, 15 minutes from an idea to a piece of software that can spend money on your behalf.
[00:05:34] That's, that's pretty good. That's astonishing, right? And even if you look at our Janus platform, 15 minutes, we can go from idea to a production app.
[00:05:47] It's. It's incredible.
[00:05:50] And, you know, you don't hear enough of these stories, but, you know, usually when you have the good part, there might be a little bit of a bad part. And sometimes it may just be in a footnote. And so it's a little bit quieter sometimes, but it shows up in many of the same reports where most of the deployments are failing. Right?
[00:06:15] Companies are having a hard time going from demo to a production type activity, from trustworthiness and scale to everything in between.
[00:06:28] And so one analyst this week found that the difference between the agent programs that scale and the roughly 74% that get rolled back is really being driven by governance.
[00:06:45] Not model quality, not compute, but pure governance.
[00:06:49] And what is governance?
[00:06:51] It's a chain of command.
[00:06:53] And ugly though, we. Well, we'll get to the ugly in full, but because I want to make sure that it has its due justice. But, you know, from now I'll just say this, okay? The same property that makes an agent useful, that it can take and act on instructions written in plain language, is the same property an attacker can use to make it act on their instructions instead.
[00:07:24] And the Audit log won't always tell you the difference. That's the problem.
[00:07:29] So that's what we're going to cover this hour.
[00:07:32] First, the good agents with wallets, what we can actually gain from them.
[00:07:38] And you know, first cracks at what it was able to gain. Then the ugly, you know, the, the broken chain and how authorized action ends up with no legitimate author.
[00:07:51] And the layer of industry is racing to build to, to, to fix that.
[00:07:58] And then we're going to close where we always close, right on the human at the end of that chain and what they're supposed to do.
[00:08:08] So as we dive into these things, I want you to think about those. I want you to think about the situations that you have with your business.
[00:08:17] I want you to think about the scenarios potentially where maybe you are a user of some of these tools or, you know, you know, trying to embed these types of softwares into your systems.
[00:08:30] Which one is driving you, which one is driving your decisions? What sort of things have you thought about as you have worked through these problems or even as you have tried to scale them in your own enterprise?
[00:08:45] Has it been a good experience, has it been a bad experience, or has it been a really ugly experience?
[00:08:52] Because I get emails and I got a whole bunch of them this week that it's really across the board. But I'll tell you, most of them have been either on the bad side or the ugly side. And you know, the few that I got to respond to this week, it was generally around trying to assist them with getting their governance in place, accountability in place, making sure folks are doing the, the right thing with those as they're pulling them into their enterprise and how they're looking at them and how they're tracking them. So we're going to deep dive into that when we come back. So I want you to stay with me.
[00:09:33] This is the most important operational conversation in AI right now and it has almost nothing to do with how smart the model is.
[00:09:45] So I think you're going to learn some things if you, if you come back. So stay with us. We'll be right back after a few messages from our sponsors. Sam, Welcome back to AI Today. I'm your host, Dr. Alan Mado. And we ended the last segment by talking about, you know, things we were going to cover this week. And, you know, we said we were going to start with the good. And what I should have actually said was we're going to start with the good, but we're going to point out a few cracks in the, the goods armor, right?
[00:10:50] And because, again, I think it's important to see how these things can transition from great and good to maybe there's a problem.
[00:11:01] So let's. Okay, so let's go ahead and just start with the good, right? I want to.
[00:11:07] I want to be, you know, generous about it, because the good, quite honestly, is why any of this matters.
[00:11:16] So two years ago, the Frontier really was a model that could answer, right?
[00:11:26] You asked, it responded, you did the work, right?
[00:11:32] Now this year, the Frontier is a model that can act, that takes a goal and executes it across, know, hours, across tools, across systems, without you in the loop, potentially for every step. And, you know, quite honestly, the.
[00:11:52] The leap is not incremental. Okay? It's really a change in. In kind. Okay.
[00:12:01] Now, whether that's large language models or, you know, how we're using and creating our agents so that they are deterministic and not probabilistic, we're gonna, we're gonna put those together, though, for this discussion just so we can just, so we can not muddy the waters too much and just really just focus on that good piece where, you know, independent of the methodology that's used, you have it going through a different set of stages. Each stage is looking at, you know, how long can it act, what tools is it using, what systems is it bringing in, and where is a human in that loop? And so that's what we're going to refine or, you know, refine it to, and we'll keep it there because, you know, if you look at how, you know, things are actually being used inside of the companies building it, it's all over the place.
[00:13:01] So, you know, OpenAI just published, you know, their own numbers this week and how they're using it. And their heaviest internal users now generate more than 60 hours of agent work in a single day. That's pretty impressive. Okay, 60 hours. And that's from one person because the agents, of course, are running in parallel while the human orchestrates.
[00:13:29] That's. That's pretty good. Ours are a little bit higher than that with Janice.
[00:13:35] We're. We're generally around the, you know, 70 to 80, because our, our methods are different and how they're orchestrated are different. But, you know, we'll. We'll say that, you know, it's about 10%.
[00:13:49] A little, you know, 10% higher than that. Not bad. Okay.
[00:13:53] But, you know, across a sample of individual users, you know, I think it was about a quarter of them had given the agent just, you know, a single task, you know, estimated at more than eight hours of human work, and that's just one instruction.
[00:14:14] So a full day's labor is being done autonomously.
[00:14:20] That's good.
[00:14:22] And you know, that is a genuine multiplier on what a person can do.
[00:14:29] You have increased their capacity to do work by a hundred percent.
[00:14:35] And I don't want to hand wave that because that's just the reality.
[00:14:41] And so if you think about it, that's 60 plus hours as a base that agents are doing work per day, and that's one in four users that are doing that.
[00:14:55] That's an eight hour task running autonomously. The math adds up pretty quickly. And so you can see where the benefits can come into play.
[00:15:05] Now, of course, we all know that doesn't happen everywhere.
[00:15:10] There are a variety of reasons for that and we're going to get into those. But now I want you to bring it into a world you and I know very well and we're going to talk about the federal world here.
[00:15:22] So this week the Pentagon confirmed it's piloting agents to automate its ATO process. It's about time. I talked about that for 12 years, all the way back to even when I was doing work for some civilian agencies.
[00:15:39] There's no reason that this process could not have been automated, even with rpa. But that's a different discussion. But you know, if you've ever lived through an ATL process, I'm going to translate it for you.
[00:15:52] And it's a security accreditation that, you know, stands between a system and a software's permission to run on that system or a user to run an application on that system. And it can, it can take you know, up to two years of documentation and processes to go through and scans to do and all the things in between to get approval to run on those.
[00:16:22] Yeah, two years.
[00:16:24] And so by the time your software gets on there, you've had to freeze it. It's two years old.
[00:16:32] You're already two years behind in your development of that application. Right. And so you can see what the problem is, you know, is with that. Now finally, they're using agents to compress that. Janice and Atlas on our end can do continuous compliance monitoring built in, just because of the things I had to go through in getting ATOs. And it's a, it's a long process, but you know, to have machines assemble the evidence, to have agents do that, to draft the packages, to, you know, draft the paperwork that's required in a specific format, that's, that's fantastic. And, but here's the part I want you to, to really understand, because it's the whole thesis of, of this show sitting inside of a government pilot, okay?
[00:17:26] They built in clear escalation paths for humans, which is mandated, right?
[00:17:36] The agent does the labor, human keeps the authority.
[00:17:41] That is the chain of command that, that we want, right? It's done, and it's done right in the, in the one place that, quite honestly, cannot afford to get it wrong. We always used to joke that, you know, you get one of these wrong and ato or not, you're going to wind up on the front page of the Washington Post. And we didn't want to be the first AI company to do that.
[00:18:05] That's why we built in these escalation procedures and these processes, and that's why security and human in the loop are everything that we focus on today.
[00:18:14] So, you know, that's the good part. And it's very substantial because the number of systems, the number of applications, you can see how much that would add up on the federal side. And that is a real capability. That's real economic weight. And, you know, it's, it's, it's real accountability that's designed from the start.
[00:18:38] Now, if that was a whole story, then this would be a short show, right? And we wouldn't have to go any farther. But I want you to watch what happens, though, the moment the humans step farther out of the loop.
[00:18:56] And here's a cautionary tale that has, has made the rounds. And, you know, it's a little funny until you, you sit with it and stuff, but, you know, the Washington Journal, the Wall Street Journal, I'm sorry, ran an experiment with an AI agent, you know, minding a small store, okay?
[00:19:19] Staff talked to, it worked. It manipulated it in playing conversation, and they got it to give away a PlayStation 5 and a pile of other goods, right? For nothing. That's great. Not by hacking it, but persuading it gently, you know, working with it and using different prompts and different discussion points, right? They didn't hack it. They talked to it. They talked it into giving something away.
[00:19:59] So sit with that for a second and think about how that's really unsettling.
[00:20:06] The agent did exactly what an agent is supposed to do. It responded helpfully to humans in front of it.
[00:20:17] Had no way of knowing that be helpful or protect the till or whatever that was, could potentially be in conflict. Had no authority structure telling it which one wins.
[00:20:33] There's no chain of command, just a capable, willing system doing what the last persuasive voice told it to do.
[00:20:48] Now scale that up.
[00:20:52] Think about a vending machine or a Treasury function. Right now you're talking about a PlayStation to maybe a wire transfer.
[00:21:08] Boy, that hurts, doesn't it?
[00:21:10] And that's the crack that can quickly be exposed under the good.
[00:21:18] And unfortunately, the numbers say that the crack is the norm, it's not the exception.
[00:21:27] I told you that 74% of these deployments get rolled back without governance.
[00:21:36] Let me add the ones that can go with that, right?
[00:21:41] You know, whether it's a Cisco security review this year that found out that 29% of organizations are actually ready to secure their agents, and one study found that almost 78% of agent deployments are out there executing high risk actions.
[00:22:05] Real tools, real consequences, right?
[00:22:09] With no deterministic policy enforcement underneath them, no hard rule that says this action you may not take, no matter who asks.
[00:22:22] That's a problem.
[00:22:24] That's scary.
[00:22:27] Now go back and ask yourself, have you interacted with these? Have you tried these? What's your experience?
[00:22:35] Maybe it's. Maybe it's reminding you of some things.
[00:22:38] So here's the shape, though, that I want you to continue to think about, though. Because we have deployed a workforce that can act at machine speed with real authority. And most of that workforce is operating without one thing. Every human institution figured out centuries ago a rule that cannot be talked out of.
[00:23:02] The good gave us the agents to do that.
[00:23:06] The bad is that we handed them most of the keys before we installed the locks.
[00:23:12] That's a problem. And that's a big gap.
[00:23:15] Really big gap. And so because of a gap like that doesn't stay empty for very long, someone is always going to move in.
[00:23:26] And when we come back, we're going to show you how the chain's broken, How a perfectly authorized action ends up with no author at all. So stay with us. We'll be right back.
[00:23:46] Sa.
[00:24:12] Welcome back to AI Today. I'm your host, Dr. Alan Mado. And last segment, we talked about the good.
[00:24:17] And we exposed a couple little cracks though, didn't we?
[00:24:21] Well, this segment, we're going to talk about the ugly.
[00:24:25] And I'm not going to soften it because you cannot defend against a threat someone's been polite about. You just can't.
[00:24:35] So start with the number that should reset how you think about this, okay?
[00:24:43] And it's. It's not great.
[00:24:46] So across surveyed organizations running agents. Okay. And from small businesses to large, I don't have the breakdown of what that survey distribution was, but 88% reported a security incident tied to those agents in the past year.
[00:25:07] Yes, 88%.
[00:25:10] That's not a risk on the horizon. That's the weather today. Okay.
[00:25:15] And against it, the same survey found something absolutely absurd, right? That on average, something like, you know, 6% of security budgets is actually pointed at the agents.
[00:25:38] Yeah, 6%.
[00:25:40] So the exposure is nearly universal.
[00:25:47] The defense around it, it's a rounding error.
[00:25:54] Yes, I'm going to say that one more time. 88% had a security incident.
[00:26:05] 6% of the budget is to defend him.
[00:26:11] So how does the chain actually break? Okay, it breaks in three ways.
[00:26:16] And let me walk each one because I think it'll help everybody understand they're not abstract. Every one of them has a real incident behind it this year.
[00:26:31] So the first one, the whole field now ranks as a number one threat. If you think about it, it's called prompt injection. Okay.
[00:26:41] And the idea is almost embarrassingly simple, something I've been doing for a long time with these.
[00:26:49] You know, an AI agent reads text.
[00:26:52] That's his whole job, but it can't reliably tell the difference between text, that's data to be processed, and text, that's a command to be obeyed.
[00:27:07] To the model, it's just all words in the same stream.
[00:27:13] That's why more complex these models become, the more difficult the prompts are and the more careful you have to be.
[00:27:24] Now, if an attacker can slip words in something, the agent reads an email, a document, a web, a tools output. You know, the agent may simply do what those words say and, you know, like, this is not a theoretical case.
[00:27:46] Happens all the time.
[00:27:48] There was a vulnerability in Microsoft's copilot, right? The zero click, meaning that the victim did nothing wrong at all. An attacker sent one really well crafted email with instructions inside of it.
[00:28:05] When the assistant did the ordinary thing and summarize the inbox, it read those hidden instructions, it obeyed them. It pulled data out of the company's own files and sent it out through a channel the company trusted.
[00:28:17] No malware, no stolen password exploit was written in English, and IT scored a 9.3 out of 10 out of severity.
[00:28:32] The most deployed AI product in the enterprise turned against its owner based on a paragraph.
[00:28:44] Here's why, though.
[00:28:47] This one keeps me, keeps me up, because really, it's really at the heart of, you know, tonight's show.
[00:28:57] When that agent exfiltrated that data, it used its own legitimate access, its own authorized tools.
[00:29:11] If you pulled the audit log afterward, you'd see an approved agent performing an approved action with valid credentials.
[00:29:21] Everything about it looks authorized. And yet no one and the company Authorized it.
[00:29:28] The order came from a stranger buried in an email.
[00:29:32] That's the broken chain.
[00:29:35] Not a link that's missing, a link that's been quietly replaced.
[00:29:40] The action still traces back to a source.
[00:29:44] The source just isn't yours.
[00:29:47] And this is the thing our traditional security teams were never organized to handle. We built decades of tools to stop stolen credentials and malicious code. And we've, you know, almost nothing stops a legitimate agent from being persuaded if it is probabilistic, especially if it's the wrong person, because it can't tell. There's no guardrails, there's no guidelines associated with that.
[00:30:25] Now, the second way you can break a chain is in the supply chain, and this one is uglier because it scales. There's no way around it. Agents are built on shared open source parts, shared marketplaces of skills and tools. MCP is a good example, right? The ease of using it.
[00:30:49] Poison one part and you poison everyone downstream.
[00:30:56] This spring, a wildly, you know, great, widely used component that routes requests for dozens of agent frameworks compromised in a public response, you know, repository for about three hours.
[00:31:15] Three hours in that window, it was downloaded about 47,000 times.
[00:31:24] Everyone pulled it in and they happened to pull in an autonomous attack bot riding inside of it.
[00:31:35] And in one popular skills marketplace, researchers found hundreds of malicious skills that were uploaded.
[00:31:47] Capabilities that look helpful, but they act really hostile just waiting for an agent to install them.
[00:31:59] That's scary.
[00:32:00] And I'm going to say this again, right?
[00:32:03] One component, live for only three hours, 47,000 downloads, had an attack bot in it.
[00:32:16] The third way, though, is, is really the simplest and it's the one you can actually feel in your gut. There's no other way to put it. And that's over permissioning.
[00:32:30] We give agents far more access than the task requires. Why? Because it's easy. Every file, every system, every tool. Why are we going to bother anywhere else?
[00:32:41] It's easier than scoping it the right way, right?
[00:32:46] A security researcher named Simon Willison gave the danger a name, and I think it's a great name and it's really stuck. And that's the lethal trifecta.
[00:32:58] An agent that has access to your private data and takes in untrusted outputs, you know, is input and can send the information back out. Just that simple.
[00:33:11] Any system with all three is exploitable, full stop.
[00:33:19] And most agents as deployed have all three because all three are convenient and they're easy and they don't take a lot of work to do it.
[00:33:30] That's the ugly injection. You can't easily detect supply chains. You don't fully see permissions granted for convenience.
[00:33:42] And underneath all of it, agentic incidences that are no longer hypothetical. They are unauthorized cryptocurrency transfers. They're data leaks that expose customer records. They're intellectual property that just happens to go out and show up on the Internet.
[00:34:02] And a new category security teams didn't used to have to think about onboarding legal liability events, right? But when an agent you deploy does harm, on whose authority does it act?
[00:34:23] That question is, is now being asked in front of lawyers. We talked about this a year ago.
[00:34:30] We knew it was coming, but we didn't know it was going to come quite as fast as it has.
[00:34:36] So if that were the end, I'd be telling you to unplug everything, right? But I'm not, not going to do that because the more useful thing happened at the same time.
[00:34:51] This was also the week that industry started building the missing links.
[00:34:58] And it's been in the open and it's been fast. And folks like myself have been leading the charge on the deterministic importance of these types of methodologies and the human in a loop pieces of these frameworks that have to be in there.
[00:35:16] But you know, one of the clearer examples really, really from a technology and process perspective was though it hit home on, you know, a couple weeks ago on the 25th and that's, you know, a company called Proof shipped an open protocol and they call it the X401. And it does one specific thing and it's of course it's overdue. We've really needed it for a long time.
[00:35:42] It lets us service, demand proof of authorization before an agent can act.
[00:35:51] Now I've talked about our ability to detect agents and orchestrate them, but we're doing it a little bit different. We're, we're going through some different processes to understand what they are and block them. The nice thing about this is it really, you know, helps that before an agent can buy anything, can sign anything, publish, move money, whatever that is, it has to present evidence of who authorized it, who's its boss.
[00:36:23] So when you read it that way, in essence asking for its permission slip, where's its hall pass?
[00:36:32] You know, it's a hall pass for software that is something that is exceptional.
[00:36:42] And you know, we've talked about our guardrail layers, right? And we've done a lot of work around that. And from bias to hallucinations to everything in between, you know, our guardrails and the deterministic policies that we're using and are fundamentally embedded in every single agent. And that means that it's a rule that our agents have to follow that no paragraph or no prompt can dictate to our agents.
[00:37:18] And that is what allows us really to make sure that, you know, our agents are doing exactly what they want or exactly what we want them to do based on the instructions that we provide it and the goals that it has.
[00:37:35] So it cannot go outside of those boundaries.
[00:37:38] Now that's your ugly, okay?
[00:37:44] And there's your, your answer. Really, side by side, the chain broke because folks have been shipping the power with the agents before the provenance.
[00:37:56] And the fix isn't slower agents, that's not, that's not going to be helpful. It's agents whose every action carries its all pass.
[00:38:06] Who it is, who ordered it, who answers for it, who's accountable.
[00:38:14] The technology to do that exists. Right?
[00:38:18] You know, we've been talking about it for a long time.
[00:38:21] The only open question though is whether folks are going to install it before the next orphaned action takes place or after.
[00:38:35] But when we come back, the one link in the chain that no protocol can build for you, we're going to deep dive into that. So stay with us. We'll be right back.
[00:39:18] Welcome back to Today. I'm your host, Dr. Alan Badot. And let me gather the whole hour in a one idea because underneath the good, the bad, the ugly, there's really only one story.
[00:39:38] The good agents can act. Now days work from a sentence.
[00:39:45] A hundred billion dollars of belief behind that, right? The bad, most of them can act without governance. It's like the wild west.
[00:39:55] They get pulled back and are not successful. They don't scale, make poor decisions, those things. The ugly, it's really the very openness that makes them useful.
[00:40:10] Lets a stranger's words become their orders.
[00:40:17] And the audit log is says everything's authorized.
[00:40:22] Three different headlines, one route, okay. And every one of them is a question about the chain of command, about whether you can trace an action back to a human who meant it.
[00:40:35] That simple. Accountability, right?
[00:40:39] Everything comes back to that.
[00:40:41] So, you know, regular viewers know the line I keep returning to on this show, okay? It's the whole way I think about these systems, how we have built our systems and why it's so important to pay attention to these cognitive agents on their own.
[00:40:57] The reasoning and the synthesis, they own it.
[00:41:02] Humans have to retain accountability for their actions. Just that simple. Just like any other employee in a boss relationship, the machine can think human answers for what gets done.
[00:41:14] This week put that sentence under Load. And guess what? It held up because it showed us the cost of the second half.
[00:41:23] Because accountability is not a feeling, it's a structure.
[00:41:29] It only exists if you can actually trace an action through every darn link that there is back to a person and an agent that acts with no traceable author.
[00:41:43] It's isn't, you know, autonomous, it's an orphan.
[00:41:50] And an orphaned action is the most dangerous thing in your entire enterprise, not because it's powerful, but because no one is standing behind it.
[00:42:02] Okay, so here's the. Here's the discipline I build by. Okay, three links. It's always three, right?
[00:42:13] You get all three, and you don't have a. A chain. You have a rope with a gap in it.
[00:42:22] The first link's identity. Okay, who is the agent?
[00:42:27] Not an AI, which one? Or, you know, acting for whom, with what scope.
[00:42:33] You know, an agent you can't name, one that you can't govern. Okay, second link is authority.
[00:42:40] By whose orders did it act?
[00:42:43] Can it prove that at that moment, when it is supposed to execute a task, it's supposed to.
[00:42:52] Before the action takes place, whether it's before the money is supposed to move or the. Before the action is approved, not after.
[00:43:02] That's the link in the whole industry they're finally catching up to.
[00:43:09] And quite honestly, it's the link that most deployments are still missing. There's a lot of work to be done around that.
[00:43:17] We know every single action. We can put a human in the workflow. We can generate a package that they have to approve. Every single piece of the puzzle is solved all deterministically.
[00:43:27] There's a reason we did it that way, to prevent these exact things from happening.
[00:43:33] And the third link? Accountability.
[00:43:36] It's always the most important link.
[00:43:39] When it's done right or wrong, the trail leads to a human who owns the outcome. Not a vendor, not a model, a person.
[00:43:51] That link cannot be bought, cannot be automated, and cannot be delegated to the thing that took the action.
[00:44:03] It's the one link you have to be in.
[00:44:10] So we spent this hour from wallets and protocols and exploits.
[00:44:18] But if you strip away all that, fundamentally, here's what's left.
[00:44:24] And it's the oldest lesson in leadership, but it's. Maybe it's got some new clothes on.
[00:44:31] Power without a chain of command is not strength, it's exposure.
[00:44:38] A capable agent with no one behind it is not an asset, it's a liability. Wearing a uniform, it's disguised.
[00:44:49] It's an insider threat.
[00:44:53] The machines, they're ready to act.
[00:44:57] The only question that matters is really the one that we started with, and that's who is able to answer instantly for everything that is run.
[00:45:17] Every single agent that takes action.
[00:45:21] How's it orchestrated?
[00:45:23] How does the Orchestrator get its instructions?
[00:45:26] What's the chain of command look like?
[00:45:30] Who authorized it?
[00:45:32] If you can walk that chain link by link to a person who will stand behind the answer that it provides, you're building it the right way.
[00:45:46] That's great. Good for you.
[00:45:48] I have no doubt you'll be successful in whatever you're trying to do with those, but if you can't, you do not have an agent.
[00:45:58] You have an orphan with your credentials.
[00:46:03] Every single system that you have access to, it has access to every single action it can see.
[00:46:13] Every single system that you have access to, maybe you haven't used in a while.
[00:46:18] I bet you the agent's going to use them pretty soon because they're going to go out, they're going to need information, they're going to grab information, and they're going to pull that in.
[00:46:29] When they pull that in, they are going to take actions that you never thought were going to be possible.
[00:46:38] Who stands behind those actions? Who is held accountable for those actions that are taken by that agent right now, today, Whether it's legal, whether it's the medical field, financial, supply chain, manufacturing, it doesn't matter.
[00:47:00] 78% are out there roaming around with no accountability and no action. Nobody standing behind those 78%.
[00:47:12] That's the number I'm going to leave you with because don't continue to build on that number.
[00:47:21] Do it the right way, Stand behind, ask questions, make sure it's giving the right answers.
[00:47:30] Be held accountable.
[00:47:34] I'm Dr. Alan Badot. This has been AI today.
[00:47:38] Give every action an author.
[00:47:42] And when you do that, you're going to be successful.
[00:47:44] I look forward to the emails that, that you all send thanking me for that.
[00:47:49] That's our show. We'll see you next.
