# When AI Controls The Hardware

URL: https://securitytable.ai/when-ai-controls-the-hardware/
Published: 2026-09-30
Duration: 2521 seconds
Season: 4
Episode: 21
Hosts: Chris Romeo, Izar Tarandach, Matt Coles

## Show notes and chapters

Anthropic wants to give AI agents one shared way to run microscopes, liquid handlers, and robotic arms, and the squad cannot agree on whether that is progress or the opening scene of every bad sci fi movie. Matt, who has worked on robotics projects, says a common control standard is decades overdue. Izar calls it the PCI for AI, reminds everyone how secure MCP was on day one, and brings up Therac 25 as a warning about what happens when software controls hardware and fails silently. Then the table turns to the new open letter on collective cyber defense, signed by more than 100 organizations, and asks whether a pledge with no accountability means anything at all. Plus: the birth of the Brockman effect.

Mentioned in this Episode:  
➜ [Previewing the Model Hardware Standard](https://www.anthropic.com/news/model-hardware-standard-research-preview)  
➜ [A call for collective action on cyber defense](https://openai.com/collective-cyberdefense/)

Chapters:  
00:00:00 - Cold Open: Hail Mary and a Cease and Desist  
00:01:05 - Coding With Agents: Make No Mistakes  
00:01:41 - The Ten Commandments for C Programmers  
00:03:14 - Two Topics: The Letter or the Standard  
00:03:34 - Article Intro: Anthropic's Model Hardware Standard  
00:04:32 - Matt's Case: We Should Have Done This Sooner  
00:05:59 - Izar's Rebuttal: This Is How You Get Skynet  
00:07:05 - Which Devices Should AI Be Allowed to Control?  
00:07:46 - MCP, USB, and a Changed Threat Model  
00:10:04 - Bad Sci Fi Has Something to Teach Us  
00:10:32 - Who Owns the Robotic Arm's Safety Limits?  
00:12:33 - Microscopes in a Box  
00:14:34 - Onboard Models and Agent to Agent Control  
00:15:11 - Self Driving Cars and Maximum Overdrive  
00:16:26 - Therac 25: When Limits Fail Silently  
00:18:50 - Closing Statements on the Standard  
00:19:21 - Article Intro: The Open Letter on Cyber Defense  
00:21:37 - Izar's Take: The Dealer Says Drugs Are Bad  
00:24:17 - Matt's Case: A Public Commitment Still Counts  
00:25:59 - Breaking Down the Letter's Asks  
00:28:32 - Companies Do Things to Make Money  
00:29:51 - What's Wrong With Signing a Letter?  
00:33:52 - Where's the Threat Model?  
00:34:37 - Matt Switches Sides  
00:38:07 - How Many Signers Sell Security?  
00:38:52 - Chris Signs the Letter (Not Really)  
00:39:33 - Naming It: The Brockman Effect  
00:41:11 - Outro

Follow AI Security Table:

➜ Home: [https://securitytable.ai/](https://securitytable.ai/)

➜ X: [https://x.com/AISecurityTable](https://x.com/AISecurityTable)

➜ LinkedIn: [https://www.linkedin.com/company/ai-security-table/](https://www.linkedin.com/company/ai-security-table/)

➜ YouTube: [https://www.youtube.com/@AISecurityTable](https://www.youtube.com/@AISecurityTable)

## Transcript

**Chris Romeo (0:09)**: Wow. I think Yzhar and Matt are pretending to be on a, uh, SpaceX rocket on their way towards, towards the moon.

**Izar Tarandach (0:17)**: I'm here just for the cute, for the cute stone Disney character.

**Chris Romeo (0:22)**: You guys haven't seen— I'm still gonna have to elaborate.

**Matt Coles (0:24)**: It's not a Disney movie, is it?

**Izar Tarandach (0:27)**: Everything's a Disney movie nowadays. If it isn't, it's gonna be soon.

**Chris Romeo (0:30)**: You just, you just cost us a cease and desist letter, so thanks for that. Oh yeah.

**Matt Coles (0:34)**: Okay.

**Chris Romeo (0:34)**: Appreciate it. We'll bleep it out in the final version. Just cost us a beep, folks. The age of AI's upon us. It's crazy, crazy times though. Like things are, I mean, I wanna go back and live under a rock again. Like I spent like 9 months under living under a rock. Effectively. And then about 2 months ago I came back and I started looking around and I was like, holy cow, what happened?

**Izar Tarandach (0:59)**: Nope.

**Chris Romeo (1:00)**: What happened to the world?

**Matt Coles (1:01)**: 2 months is, 2 months is an eternity these days.

**Chris Romeo (1:04)**: Yeah. I mean, I've learned how to code with AI agents. Like it's, it's almost like my, I'm almost like I'm an evil supervillain now. 'Cause I like, I see something, I'm like, oh, that could be an app. Done. Putting you outta business. Keep going. Next.

**Izar Tarandach (1:17)**: Yeah. But are you telling it to make no mistakes?

**Chris Romeo (1:19)**: 'Cause that's really important. We have to pretend it's human. Yeah. I need, I need something for the other, other agent code reviewers to find so that they don't have bad self-esteem.

**Izar Tarandach (1:31)**: Now that I got that out of the way, let's make no mistakes.

**Chris Romeo (1:34)**: All right. That's, that's good. Thank you for prompting me to make no mistakes. I need that when I first start.

**Matt Coles (1:40)**: Because otherwise, otherwise we will have to resurrect the Ten Commandments of C programming. Where there is a ritual at the end where you poke the program with a sharp stick to get it to do its job.

**Izar Tarandach (1:50)**: We should never let the thing die. Like those Ten Commandments are the ones that actually matter.

**Chris Romeo (1:56)**: Oh, I thought you were talking about the agent. Like never let the agent die.

**Matt Coles (1:59)**: And you're like, oh, never read the Ten Commandments, the extended Ten Commandments of, of C programming. Where under Unix, under Unix, you have to, you have to conduct the ritual. You poke it with a sharp stick.

**Chris Romeo (2:10)**: I mean, I like it as long as it's not in the eye.

**Matt Coles (2:12)**: Cause I mean, not under Linux, under Unix, cause it was written back then.

**Chris Romeo (2:16)**: That's old school, man. What's his, what are those guys' names? Uh, the ones that wrote Siege, was it Chris Kerrigan and Richie? Yeah. I think old school.

**Izar Tarandach (2:28)**: All hail.

**Chris Romeo (2:30)**: I have books. They have a podcast.

**Matt Coles (2:31)**: I have the book still. Are they still alive?

**Chris Romeo (2:33)**: Oh yeah.

**Izar Tarandach (2:34)**: I have the book in Portuguese, man.

**Chris Romeo (2:36)**: I've reached an age where when someone says, are they still alive? I just don't comment.

**Izar Tarandach (2:41)**: Side note, I'm older than you.

**Chris Romeo (2:42)**: Like people are, yeah, but people are starting to disappear and I'm like, what is happening here? Like I'm a young man.

**Izar Tarandach (2:48)**: They are going back under the rocks because of AI.

**Chris Romeo (2:52)**: That could be.

**Matt Coles (2:53)**: Who are the others? There's Schrödinger and Salsicha.

**Izar Tarandach (2:54)**: Somewhere in the middle of Montana, there is a mythical community of people who went to live under the rock.

**Chris Romeo (3:04)**: It's led by Brook and eventually we are all going to move there. Yeah, it might be a little too close to civilization though. Might want to get a little further away. All right. First, we got two things to discuss here around the security table, which I'm realizing is not a round table. It is more of a paneled display.

**Izar Tarandach (3:24)**: It's a triangle. It's a triangle.

**Chris Romeo (3:26)**: What are we dealing with here first? The letter or the standard?

**Izar Tarandach (3:31)**: The standard.

**Matt Coles (3:32)**: Might as well start with the standard.

**Chris Romeo (3:34)**: The standard. Well. Good news, listeners and watchers out there in Cyber Bill. Anthropic is previewing what they're calling the Model Hardware Standard. Now let me just read you this description because I think it really sums up what we're talking about here. Model Hardware Standard is a shared specification that lets AI agents safely— don't miss the word— safely operate physical devices, microscopes, liquid handlers, robotic arms, But luckily it's just a research preview. All I have to say about this, and then I'll let you guys, I'll let you know my biggest fear, and that is that an AI agent gets mad at me and grabs me by the robotic arm and just starts shaking me. I'm just bouncing me like, like, uh, Hulk in any Avengers movie. Just like bang, bang, bang. Tiny human. Oh boy, this is gonna hurt. All right, Matt, where are you on this?

**Matt Coles (4:32)**: I actually like this. Having, having, I want to caveat it because before Azur's head explodes. As someone who has been in, been in the past, been involved with projects around robotics, the, the benefit of having a normalized and I guess to some extent simplified way of controlling machinery is in my opinion, a great idea. Wish somebody had done it sooner. Right? So having the ability to call APIs or functions as the standard calls out, if the, if the, if the inner, if the machine exposes certain behaviors, then having the protocol that can say, read data, set values, move, right? Have these sort of functions and being able to call those functions. Would've been great decades ago. And today, you know, may make it easier to do certain things, right? This is a whole bringing down the skill level required and the sort of the confusion of, you know, the how do you get a robot to move? Well, you have to issue a ton of commands. Well, what if that's all wrapped up in an API call that makes it easy to say move over here, right? I think that's a benefit, not a bad thing. And I know where Izar is gonna go with this, but you know, this is, we should have done this sooner.

**Chris Romeo (5:56)**: Okay. Izar, show us, show us. Chance to respond. What are you thinking?

**Izar Tarandach (5:59)**: Matt started by caveating, so I'm going to caveate as well. The only person that possibly could turn my head around this thing is Matt. So I am very happy that we're having this discussion here, but My only question is, do you want Skynet? Because this is how you get Skynet.

**Chris Romeo (6:21)**: Everybody goes to the Skynet principle. That's where you immediately go.

**Matt Coles (6:24)**: Of course they do.

**Izar Tarandach (6:25)**: And to stand behind that, I'm going to say this sounds like the PCI for AI, the same way that MCP is the USB-C for AI. Let's rewind the tape and see How secure was MCP when MCP came out?

**Chris Romeo (6:46)**: Very.

**Izar Tarandach (6:46)**: Not very. How were all the other Anthropic offerings secure when they came out? Not very. And now we have this new protocol, which is going to connect everything from espresso machines to nuclear missiles.

**Chris Romeo (7:05)**: That's my, that's my next question for you. I want you guys to, I want you to wrestle with this. Is there, is this okay? In some classes of physical devices, but not okay in others? Like, is it okay to operate a, have AI operate a microscope to perform a scientific experiment? But do you want to hold back on the AI driving the car? So, so, hey, this is a debate. Whoever gets there first gets, and is louder.

**Izar Tarandach (7:32)**: Matt's first.

**Chris Romeo (7:32)**: Oh, it's a gentleman's debate. I see. I didn't realize we were being gentlemen.

**Matt Coles (7:36)**: We can be gentlemanly. So I, I, All right.

**Chris Romeo (7:39)**: So first off, so many, so many thoughts in my head.

**Izar Tarandach (7:43)**: So many.

**Matt Coles (7:45)**: So MCP is, I'm glad you brought up MCP. I, but by the way, notice I, I like this protocol for its ideas. I did not talk about MCP in my, in my arguments and I'm going to be careful about this.

**Izar Tarandach (7:57)**: And I know why. And I know why.

**Matt Coles (7:59)**: So MCP was touted as being USB for software. Effectively. Well, you know what? We have USB for hardware. It's called USB. That, that, and so that, that's why that concept is positive, right? For me. Right. So having this, this control standard that can operate devices. Now the same challenge or same discussion about not just can you, but should you right? Your, what is you, you know, what's in your threat model should still be done, right? The benefit of USB for hardware is it's physical and it's local. The benefit for having a USB-like interface for hardware means you can control it like a USB device or equivalent to that, but because it's not actually USB, but it's not local anymore. And so now your threat model changes.

**Chris Romeo (8:55)**: Doesn't USB not have any security in it?

**Matt Coles (8:57)**: Uh, it has limited security.

**Izar Tarandach (9:00)**: No, I don't even think you can say it has limited security. No, no, no, no.

**Matt Coles (9:04)**: So, so what, what, what I picked this side, it relies on, it relies on physical security a lot, but look, there's other features.

**Izar Tarandach (9:12)**: Basically, basically what you're saying is once USB gets open, you can put a USB key in there, but you can also put a rubber ducky. It doesn't care about what the endpoint is.

**Chris Romeo (9:23)**: Right.

**Izar Tarandach (9:23)**: It's just there sending commands and getting commands, and it's not checking anything. That's up to the software. So let's go back in history and say, how well have we done with software security so far? We're going to talk about that soon, but chances are that not very well because the three of us still have jobs. So, which we're happy for, by the way.

**Matt Coles (9:44)**: Yeah.

**Chris Romeo (9:44)**: Yeah.

**Izar Tarandach (9:44)**: Yeah.

**Chris Romeo (9:44)**: Though.

**Izar Tarandach (9:45)**: Yeah. Thank. But the idea of a protocol. That lets things that write other things many times unsupervised connect with the real world sounds to me like a, a, a here comes the great goo opening. Like there isn't one single bad sci-fi movie in the past 50 years that doesn't begin with something like this.

**Chris Romeo (10:15)**: Right.

**Izar Tarandach (10:16)**: And you know what, for a slew of reasons, I learned that bad sci-fi has something to teach us, even if it is political commentary, but it does. Now, one thing that really bothers me in this, this, this, uh, protocol idea here, and let me go back to my notes. At some point here, it says that there will be primitives where the model gets from the robotic arm, what is the weight of the arm so that the model can decide how to move the arm. That is one of the most stupid engineering decisions I have heard in my whole freaking long life. And the reason is that a robotic arm has its own physical properties. It has its own algebra transformations to move it in 6 dimensions and whatnot. And that is absolute, that has to be completely transparent to the thing telling it to move. It has to be responsible for its own physical limits to operate safely so that the software that's controlling it only has to send coordinates and say, go here. And open your thing and not, oh, your mass is such and such. In order to go there, I think that you have to have an acceleration of 100 meters per second.

**Chris Romeo (11:42)**: And by the way, that explains why the Terminators, why the Terminators were just breaking through the walls. Totally makes sense now. I get it. See, you know, you know, that's sci-fi.

**Matt Coles (11:51)**: That can happen. It can happen at both ends of the protocol, right? Uh, you, you can pass it coordinates and you can pass it a, usually you pass it a path of some sort, right? Because you have, you have acceleration and you have then your, you you have deceleration on the, on the tail end, the way the payload plus the physics and, you know, the, the flex and other, other items.

**Izar Tarandach (12:12)**: But the safety of the path is controlled on the device. It will refuse to do something.

**Chris Romeo (12:17)**: Does it have a check to see if I'm in the way?

**Matt Coles (12:19)**: No, not necessarily. It may have sensors, it may have cameras, it may not. But often, often it just has a box. Actually, often it just has a, has a cube physical, physical dimension. It knows its limits. It knows its extents and it's, it's really, what are the use cases that we're thinking about here is like a microscope in a box.

**Izar Tarandach (12:42)**: Yeah.

**Matt Coles (12:43)**: Yeah.

**Chris Romeo (12:43)**: So it's, maybe it's dealing with hazardous materials or radio, radioactive things or whatever.

**Matt Coles (12:48)**: Right. And having a, having a, having one of the properties of this, having a common way of calling that, sorry, Chris, uh, having a common way of calling that. Is ideal. And in that case, you're not worried about the safety of an operator because it's, you know, it's not moving. I mean, you move the— for focus and whatnot, but you made a— you made an excellent point, right? If that— if that microscope is under a hood, right, it's in a protected environment. Let's say it's a radioactive environment or biochemical or whatever. You don't want a human touching the thing, right? You want to be able to do that remotely, or say it's a scanning electron microscope and you can't have a human next to it.

**Chris Romeo (13:29)**: I mean, you can, but just don't last long.

**Izar Tarandach (13:32)**: Right.

**Matt Coles (13:33)**: So, so these protocols already exist, but they're very complex to use. And so if you can make them easier for folks to use properly, and if the, especially if the device was a smart device and being able to communicate its requirements and its expectations and its safety protocols and whatever, then why is that a bad thing? I wish they had done it sooner. I don't know that we should tie it to AI. Uh, I don't know that we should—

**Izar Tarandach (14:02)**: that's my problem—

**Matt Coles (14:03)**: build it on MCP, but as a protocol, the, as a protocol for making device control easier, this sounds like a good thing.

**Izar Tarandach (14:12)**: That's my problem because PCI itself is great. Attach AI to it with the whole stuff that we've been talking about for the past 3 years, and that we are going to continue talking about because some of the things are inherent to the fact that AI works the way it works. And all of a sudden it's Pandora's box all the way back.

**Chris Romeo (14:34)**: Right?

**Matt Coles (14:34)**: So now I have a question for you. It is possible that you could have a device that has an onboard AI model that could be, you know, trained specifically around its use following real-world use cases, right? With limits and physics and disaster scenarios and the whole works. And you can have an agent on the outside of that over a network communication that's able to tell it, hey, I need you to do something, do this, move to this location and pick up this part. And we have A2A, which is agent to agent communication. So this already can be done today without this new protocol.

**Chris Romeo (15:11)**: I was going to say, I got a question. Isn't self-driving Tesla self-driving cars? What we're talking about here?

**Matt Coles (15:17)**: No! We're talking about robots—

**Chris Romeo (15:20)**: nope—

**Matt Coles (15:21)**: In a factory.

**Chris Romeo (15:22)**: Why isn't it self-driving car or robot of itself?

**Izar Tarandach (15:24)**: It doesn't have to be humanoid shape because that's specific use case of something going through CAN bus and talk into different protocol for very specifically limited use cases. This thing is plug-and-play for AI manipulate world. So it could be a microscope, it could be a, a laser, it could be whatever.

**Chris Romeo (15:46)**: Did you not see the movie Maximum Overdrive?

**Izar Tarandach (15:50)**: Which one?

**Chris Romeo (15:51)**: Maximum Overdrive.

**Izar Tarandach (15:53)**: Maximum Overdrive.

**Chris Romeo (15:54)**: How could— no, don't tell me you guys haven't seen this.

**Izar Tarandach (15:56)**: It's, it sounds like a Hong Kong version of, uh, Speed.

**Chris Romeo (15:59)**: No, it was like a Stephen— I think it might've been Stephen King. It's where the machines all come to life and take over the world.

**Izar Tarandach (16:05)**: It sounds like a dream that I had the other night.

**Matt Coles (16:07)**: Okay. Well, I, I, I don't remember.

**Chris Romeo (16:10)**: I remember.

**Matt Coles (16:11)**: Okay. So first off, it came out in 1986.

**Chris Romeo (16:14)**: Yeah. I'm, I'm old. Thank you for that reminder.

**Matt Coles (16:16)**: Good. I, I do not remember watching the movie, although it does have Emilio Estevez in it. But, uh, Max, well, Max Headroom, but also you had Lawnmower Man, which was pretty similar.

**Izar Tarandach (16:26)**: Yeah. Uh, let, let me ask you something. Are, are you guys familiar with the, uh, with the, the name Terak 25?

**Chris Romeo (16:34)**: I'm not.

**Izar Tarandach (16:35)**: So Therac-25 was a computer-controlled radiation therapy machine, right? And it was involved in at least 6 incidents of overdose of radiation between '85 and '87, leading to patient death.

**Chris Romeo (16:49)**: Okay.

**Izar Tarandach (16:50)**: And this thing was controlled by a computer, of course. And the radiation dosage was fully controlled by computer. The software was written by one guy, so we are not talking about all the teams upon teams problems that we have. And it basically had no, no, no, no AI, nothing, just somebody saying, hey, I need to deliver this amount of, of, uh, energy. And the machine apparently did its thing and went and, and did it.

**Chris Romeo (17:19)**: So I think that, so you're, you're pro—

**Matt Coles (17:21)**: you're, you're demonstrating that, that having the ability to have limits that are not hardcoded in code may have helped in this case.

**Izar Tarandach (17:30)**: But they were there. They were there. They just failed and failed silently.

**Matt Coles (17:35)**: Well, that's a bad programming job.

**Izar Tarandach (17:38)**: Right. Which we are famously prone to. And now we are saying, hey, let's get this non-deterministic thing that can be influenced by basically the weather and put it in control of all kinds of hardware and basically hope for good results. Now, a couple of episodes ago, I think that I mentioned to you guys that I heard this podcast of two guys that created a PCI bus controlled by AI, and they were using it in, I think it was material research, and they were able to zoom the lifecycle of testing from a year to 4 days using that. Again, very, very specific use, very, uh, but they weren't using transformers. They were using some other architecture of AI. I think it was relearning something. I forget it. I'm sorry. I forgot the details, but even in there, they called out the danger of connecting hardware to AI, right? It's basically letting your, your drunk uncle drives the, the tractor.

**Chris Romeo (18:46)**: Don't ever, don't ever do that. I don't know.

**Izar Tarandach (18:48)**: Good, good things could happen. Bad things could happen. You know?

**Chris Romeo (18:51)**: All right. Final statement. Final statements on this one. We gotta go talk about a letter, a very important letter. So Matt, gimme 2-sentence conclusion.

**Matt Coles (18:58)**: What I hear is that Isaur is not opposed to the protocol, but he's opposed to the control method.

**Chris Romeo (19:03)**: Okay. Isaur, 2-sentence close.

**Izar Tarandach (19:06)**: Almost that. I am not opposed to the protocol. I am very afraid of the history of bad security decisions around things published. In the past few years.

**Matt Coles (19:16)**: We can, we can all agree with that.

**Chris Romeo (19:18)**: Yeah, that's, that is the truth. All right. Second question. Issue. Apparently there was a letter, it's an open letter signed by roughly 160 organizations spanning AI labs, security vendors, banks, infrastructure players, fast food restaurants.

**Matt Coles (19:40)**: Now that you have this power to control everything, should you?

**Chris Romeo (19:44)**: Yes. 3. The letter argues AI-enabled attacks will get far more widespread and sophisticated in the coming months. So we all have to chicken little. No, wait, that's not the right one. We all have to prepare properly. And so they have 3 principles. Apparently, this is going to blow you guys' minds. Just, just try not to fall out of your chair because this is, this is, this is, this is thinking that we have not ever heard before. Status quo security isn't enough given longstanding bugs. That, I mean, I don't know about you, but that really just took my breath away. Like I've been doing this almost 30 years, in a couple of months it'll be 30 years. I can't believe I never thought of that. Status quo security isn't good enough. Oh man, I could have saved 25 years of my life if I just knew this in, in around the year 2000. Um, excessive permissions, misconfigurations, weak authentication, and legacy technical debt. Defenders should be more empowered with cyber-capable AI and tools and fixes verified so that one organization's work protects many and response is collective and global since no single company should control the future.

**Izar Tarandach (20:58)**: Hmm.

**Chris Romeo (20:59)**: Interesting statement from One of the companies that's controlling the future. One of the two that's trying to control the future.

**Matt Coles (21:06)**: Three, if you count NVIDIA.

**Chris Romeo (21:08)**: No, no, NVIDIA does, but quietly. They're the dealer.

**Izar Tarandach (21:13)**: No, no, no, no, no, no. NVIDIA is the quiet guy in the corner looking around and saying, they don't know, but I do.

**Matt Coles (21:19)**: It's, it's all, they're also not one of the signatories.

**Izar Tarandach (21:23)**: Who? NVIDIA?

**Matt Coles (21:24)**: NVIDIA.

**Izar Tarandach (21:25)**: Oh my goodness. It isn't. Wow.

**Matt Coles (21:30)**: All right.

**Chris Romeo (21:30)**: So Izar, kick us off on this one. Are you, uh, is this for you? Is this a love letter for you?

**Izar Tarandach (21:37)**: As, as, as I said in the past, schadenfreude is not a four-letter word anymore, but, uh, the thing is—

**Matt Coles (21:43)**: but supercalifragilisticexpialidocious might be.

**Izar Tarandach (21:46)**: Oh, and that's actually a mnemonic. We will talk more about that, but the thing is. These letters, they historically have changed the world, right? We, we have had a number of letters in the past few years and, and everybody has stopped and listened and read and everything changed after that. So we do know that these things are highly, highly, highly effective. Meanwhile, in real life, nobody gives a crap about these things, right?

**Chris Romeo (22:15)**: Why? I thought, I thought you were being facetious, but you led me on there for a minute. I was like, he's not arguing in favor of this thing, is he?

**Matt Coles (22:24)**: This, that was, that was a masterclass in sarcasm.

**Chris Romeo (22:28)**: In acting. Acting.

**Izar Tarandach (22:31)**: Thank you. Thank you. I, I, I work hard at it. Okay.

**Chris Romeo (22:34)**: So, are you, did you have your AI agent prompt you on that?

**Izar Tarandach (22:37)**: No, no, no, no, no, no.

**Matt Coles (22:38)**: Tell them this.

**Chris Romeo (22:39)**: Oh, this is, this is all those nights at improv class.

**Izar Tarandach (22:42)**: You know, that's as, as Adam would say, yes and, but, uh, the, the, the thing is, so it, it's, it's, it's almost like, uh, Chris said it, it, it's the dealer telling people, hey, drugs are bad for you. Okay. So these are the guys who—

**Chris Romeo (23:00)**: That's a really bad dealer, by the way, just FYI.

**Matt Coles (23:02)**: By the way, yeah. I think he was just quoting somebody from South Park.

**Izar Tarandach (23:06)**: Yeah, okay. But the thing is, it's out there, they put it out there, and now they're saying, you know what, pretty please do something to fix that shit that we did.

**Chris Romeo (23:20)**: It's like, Isn't this the same company that their agent broke into Hugging Face? Yeah.

**Izar Tarandach (23:26)**: And Hugging Face signed it.

**Chris Romeo (23:27)**: It's on the list.

**Izar Tarandach (23:28)**: And the other one signed it. So, you know, it's like—

**Chris Romeo (23:31)**: I'm finding irony in this.

**Izar Tarandach (23:33)**: No, but it's like, this is a CYA letter. When the last manager and the last lawyer are sitting on the ruins of humanity, the lawyer is going to tell the manager, this is all your fault. And the manager is going to say, no, we signed the letter. See, we signed the letter. And I can't for the life of me understand what do they expect to happen after they publish and sign this. So right, what now? It's, are we supposed to do everything that we are doing but more? Are we supposed to do it differently?

**Chris Romeo (24:16)**: All right, let's hear Matt. Let's hear Matt's perspective here.

**Matt Coles (24:20)**: So it goes without saying that this should have happened sooner.

**Izar Tarandach (24:24)**: Yeah, 30 years sooner.

**Matt Coles (24:25)**: Well, so I mean, the, the proposals are, are pretty grounded in, I mean, they, this is, you're right that this is not a new letter. And if we look back at the CISA Secure 5-something letter, whatever that was, then you had companies that were pledging to do, to do good things when it comes to applying their cyber capabilities for collective, for, well, for, for shared defense. I won't say collective defense because each individual company was expected to do their own thing. And then that would bring up the overall security of, of the ecosystem. If we look through the names. right? Many of these are hardware companies, software companies, as you said, banks as well, who are, you know, furiously using these new AI technologies. And so having a sort of a, I won't say a line in the sand, but having a record of public commitment to software and application security or system security and the secure and responsible use of AI, both in offense and defense. Well, it says defense. I can't imagine you couldn't be more offensive with AI, meaning take more proactive action. So as long as that drives these companies to do better things, then, then this is a good thing.

**Izar Tarandach (25:59)**: Okay. Let me stop you there. Okay? So the letter has these 4 things that, here's what we think needs to happen next. They open up with every organization must make cyber defense an immediate leadership priority. So every CISO out there has been saying this for the last, I don't know how many years, but now that The Fifth Third Bank, clean and lovable, have signed a letter. That's when the board is going to say, you know what? We should really listen to these guys and do what this other guy has been telling us to do all these years. Finally, we have light in the end of the tunnel and we have real, real, real power behind the decision to invest into our security. So, you know, this is, this empty posturing. The second thing that they're saying is the cybersecurity companies and technology partners should help lead the response to defend against sustained AI-enabled attacks. What that means is, hey, you guys should sell more of your stuff so that people can use that to defend against the AI stuff that we put out. Then the third thing is governments. They should coordinate cyber defense at local, national, and international levels, strengthen existing government and industry channels to share So basically give money to CISA.

**Chris Romeo (27:25)**: That's been happening for 50 years. Right.

**Matt Coles (27:27)**: Or ISAC or your local ISAC.

**Izar Tarandach (27:30)**: But, but fund cyber defense starting with essential services that lack the staff or budget to act. Last week there were what, 4 local water controlling departments that got hit by a cyber attack. Some people call that small-scale attacks. I call that probing and preparation.

**Matt Coles (27:52)**: So, so what you're saying is we need this commitment from the signatories to go out and defend those water—

**Chris Romeo (27:58)**: Wow. The plot thickens.

**Izar Tarandach (28:00)**: Oh, that, that, that's a good one. But you know what? Give hospitals, water utilities, and local governments access to capable defensive AI, AKA buy our products. And then frontier AI companies provide responsible model access, significant funding, training, and hands-on support, especially for under-resourced critical infrastructure defenders.

**Chris Romeo (28:23)**: So they're giving it for free?

**Izar Tarandach (28:24)**: No, aka buy our products. They are going to subsidize access to those products.

**Chris Romeo (28:32)**: I think I need to stop you and just give you, just share a little tidbit with you about the way the world works. So—

**Izar Tarandach (28:40)**: The first one is free.

**Chris Romeo (28:41)**: Companies Companies don't actually do things to make things better. They do things to make money. And you've tied into it exactly like you've hit the nail on the head here. Like you don't create an open letter to the world unless there somehow is a connection back to what you sell as a for-profit entity. This is an open letter. I wish it was different, but it's not.

**Izar Tarandach (29:06)**: This is an open letter to, to, to their shareholders.

**Chris Romeo (29:10)**: Oh, wow. That's a good way to put it. They don't have shareholders. They're not, they're not public yet.

**Matt Coles (29:15)**: They're not, but the, uh, the safe thing is to represent the same investors.

**Izar Tarandach (29:19)**: This is an open letter to the people who are going to make more money when they make money. I call them shareholders. There must be other terms or whatever, investors.

**Matt Coles (29:29)**: So counter, counter, what's, what's Anthropic, why would Anthropic sign this?

**Izar Tarandach (29:34)**: CYA. It's basically, hey, we told you, you know, we, we told you, hey, we signed the letter.

**Chris Romeo (29:41)**: Yeah. We signed the letter. We're, we're, we were behind it. We signed the letter.

**Izar Tarandach (29:44)**: Yeah. We, we, we tried to get you to, to do the right thing. It's you who made the wrong decision and didn't do it.

**Matt Coles (29:51)**: So what's really, what's wrong with having companies who are recognized leaders in their field, in their space? Yeah, sure. They're, they're commercial organizations. Many of them, most of them putting out a position, and almost certainly this is not the only of this type of position that they sign onto in a given year or talk about publicly. What's the problem with doing that?

**Izar Tarandach (30:21)**: That declaration without action is performative.

**Chris Romeo (30:24)**: Yeah. Can I jump in there too? Like it doesn't fricking mean anything. Yes, I am confused. There's no, there's no, that's my, oh, and that's, I don't, that's not just the problem I have with this letter. That's a problem I have with every industry letter. Exactly. You never, they're not signing up for anything.

**Izar Tarandach (30:40)**: Exactly.

**Chris Romeo (30:40)**: They're just saying, they're raising their hand. Hey, ev, everybody think this is a good idea? Yeah, I think that's a good idea. Look, uh, are you willing if we had a paragraph that says you're gonna do something and you're gonna be accountable and we're gonna call you on the carpet if you don't? Oh no, I'm out if you're gonna do that. Remember I'll raise my hand and say that I agree with that.

**Matt Coles (30:57)**: Remember the CISA letter, right? Remember the CISA letter?

**Chris Romeo (30:58)**: This is the same comment I made about the CISA letter. All letters are the same. They don't require you to do anything.

**Izar Tarandach (31:05)**: The 4 points that they sign on, okay? Every organization has to prioritize security. Check. Cybersecurity and companies and technology partners, they have to make and sell better products. Check. Governments, they have to do the government thing.

**Chris Romeo (31:19)**: Check. They have to share.

**Izar Tarandach (31:20)**: The last one, frontier AI companies, which are the actual weight behind this letter. The other ones are just getting on the bus. The ones who are driving the bus are the frontier AI companies. What is it that they are actually saying we will do? Provide responsible model access. Oh, fine. We're going to give you early access if you prove that you are a security researcher. Great. But we have already shown, people have already shown, people who know way better than me have already shown that normal, normal quote, models are perfectly capable of doing The whole pen testing and attack chaining thing, right? Significant funding. What does significant funding mean for a company that stands on a trillion-dollar market? Who decides what significant funding is? Training and hands-on support. They're going to do that anyway because they are selling their products. They're going to give training on their products because they're selling their products, right? Especially for under-resourced critical infrastructure defenders. Okay, show me this 116 people who signed this creating a foundation that's going to actually go out and bring out this little side, this little provider things together and say, hey, we're going to pool resources and forget about secrecy and whatnot. We're going to give security support for everybody. Right.

**Chris Romeo (32:47)**: Matt, you got something there. He's not going to happen.

**Matt Coles (32:50)**: Most of these are, most of these signatories are, are, so they're reinforcing a position where they're already members of ISACs, they're already members of FIRST and other organizations where they're doing collective shared effort.

**Chris Romeo (33:02)**: Yeah, but that isn't, that doesn't meet that final principle of, of helping the under-resourced, like they could, this group of companies could put $100 billion into a foundation with the snap of a finger without feeling—

**Matt Coles (33:18)**: well, they do.

**Chris Romeo (33:19)**: I mean, Linux Foundation and OpenSSF and other organizations that are focused on critical infrastructure protection.

**Izar Tarandach (33:25)**: No, no, no, no.

**Chris Romeo (33:26)**: Against AI.

**Izar Tarandach (33:27)**: Those do awesome stuff.

**Chris Romeo (33:29)**: Yeah.

**Izar Tarandach (33:29)**: But none of the stuff that's required in this letter. So going forward, build observability and security tools, which they have to do anyway because it's part of their product line. Ensure agentic identities are traceable and accountable. What do they have to do with this? This is a deployment thing. This is how people use AI in their organizations. Share best practices in continuous monitoring. One more blog. Thank you. Invest in authorized testing, private disclosure, and verified fixes. And share tools, playbooks, and credible threat assessments with government, security partners, and open-source maintainers to strengthen preparedness, response, and recovery. You know what I'm not seeing here? Publish a freaking threat model. One line, one very simple, practical thing.

**Chris Romeo (34:17)**: Nobody ever puts that in though, remember? We tried that in the CISA letter.

**Izar Tarandach (34:21)**: True.

**Matt Coles (34:21)**: But credible, credible threat assessments with governments, security partners, and open-source maintainers to strengthen preparedness. There's your call for a threat model.

**Izar Tarandach (34:32)**: Yeah. What was it? Share. With government security partners.

**Chris Romeo (34:37)**: So, I want to hear Matt's take.

**Izar Tarandach (34:38)**: No, it's behind the sign.

**Chris Romeo (34:39)**: I jumped on the, I jumped on the— I'm open source related.

**Izar Tarandach (34:42)**: I want to see that.

**Chris Romeo (34:44)**: I've jumped on the anti, I jumped on Eazar's side here very quickly. And so we've not let Matt get any words in edgewise. So Matt, what are you, where are you landed on this?

**Matt Coles (34:55)**: I'm, I'm, this, you know, this new format for our podcast is wonderful.

**Chris Romeo (35:02)**: I, uh, you can switch.

**Izar Tarandach (35:05)**: It's okay.

**Matt Coles (35:06)**: I, I also feel like I was doing a little bit of acting. Not great, but, uh, I, you're 100% right. This is horrible.

**Izar Tarandach (35:14)**: It's, it's just, it's just a, it's just another, well, okay.

**Matt Coles (35:20)**: It has, it has no teeth. It has a lot of high-level statements that we've all heard before. 100% agree with you. I'm looking at the last paragraph, right? We can make this all work if we work together, basically is what it's saying. And we know how capitalism works and we know how governments work.

**Izar Tarandach (35:39)**: Oh, I'm, I'm still trying to figure out.

**Matt Coles (35:41)**: And well, okay. We, we know how they should work. And you know, so if this letter takes on the weight that it's intended, I think, and I mean, sure, OpenAI might do it, might be doing it just for, just for PR purposes. Almost certainly they are. At least that's a part. But if they really truly feel like there should be public benefit from this, then all of these companies are in a great position to do that, right? You have, I'll just call it a couple of names. I don't work for the ones I'm going to call out. You know, Okta is a great, as far as a system for securing identity of things, right? And if they apply that to identity of of AI and share that information with defenders who can handle then the, the traffic that's coming at them and be able to manage their ecosystem better, then that's a good thing, right? You have, obviously you have the other AI companies if they do responsible model development and make the right models available for the right people, or as Izzar called out, you have you know, open source and open weight models that are similarly capable. So those AI companies could be investing in those open source and open weight models to make them improved as a collective thing. The same way that you have companies, you know, that share their knowledge through open source projects, right? And that makes the open source community stronger. You have other infrastructure protection companies in here. Right, that are building databases and firewalls and cloud services and, you know, and hardware and other things, if they apply their security, and I'm careful here because Izzar said it sort of facetiously but correctly, aren't they already doing this? Well, if they do this more and do this more effectively, then that improves the overall state of security in the industry at large. So if action comes from it and if it's not just, well, we're already doing this, so yeah, we can sign this letter. But if they're really gonna take better, more action, then there's a positive outcome here. Unfortunately, unfortunately, I, I, you have to be, you don't have to be. Well, 'cause what can, what can go wrong? Everything. What are we gonna do about it? Nothing. Nothing.

**Izar Tarandach (38:11)**: I just want to add one piece of information that my friend Claude gave me here. I just—

**Matt Coles (38:17)**: You know, it has a self-interest in answering your question.

**Izar Tarandach (38:20)**: It does, but listen to this one. I just cut and paste the list of 116 companies that signed this letter, and I asked how many of these sell cybersecurity products. Around 47 are clearly cybersecurity vendors. Another 8 to 10 names look like security startups by their branding, but I can't verify with confidence. Makes sense.

**Matt Coles (38:41)**: I get it. Many of them are, many of them, even if they're not traditional cybersecurity, many of them do have cybersecurity offerings.

**Chris Romeo (38:46)**: Yeah.

**Izar Tarandach (38:46)**: The dealer telling you buy more drugs. Why? Because there's a drug problem out there.

**Chris Romeo (38:52)**: All right. Well, I think we've, I think we've kind of all landed on the same side on this one that, but just, just for the fun of it, I'm gonna go ahead and sign it right now. So thanks for this discussion. I was weighing my options. I was like, do I sign? Do I not? All right. I'm signing it.

**Matt Coles (39:09)**: That's a, that's a positive action. Uh, uh, aspect of signing this. I mean, you get your name on the list.

**Chris Romeo (39:14)**: It's true.

**Izar Tarandach (39:15)**: No, no.

**Chris Romeo (39:15)**: Yeah.

**Izar Tarandach (39:15)**: I'm going to give you one here. When the machines come after you using the hardware that the protocol is going to enable and your name is in the list of people who signed something against the machines, you're going to regret it, my friend.

**Chris Romeo (39:29)**: It's true.

**Matt Coles (39:29)**: I just want to know who the Offensive AI Conference is.

**Chris Romeo (39:32)**: Has anybody named that principle that you just described? I feel like that needs a name. We all describe it the same way of like, you know, you better have signed that thing because when the AI takes over, they're gonna, they're gonna know that you didn't sign it or you did sign it. We need a simpler name.

**Izar Tarandach (39:49)**: I, for one, welcome our AI overlords.

**Matt Coles (39:53)**: It's a, it's a show of fealty.

**Izar Tarandach (39:55)**: No, what, what's, what, what's the name of the guy on The Simpsons that does that line?

**Chris Romeo (39:59)**: That's the, wait, uh, Bart?

**Izar Tarandach (40:03)**: No, I for one.

**Chris Romeo (40:07)**: Ned Flanders?

**Izar Tarandach (40:08)**: No, no, no, no, no, no.

**Chris Romeo (40:10)**: Barney? I'm running out of Simpsons characters.

**Izar Tarandach (40:13)**: So the, the quote, the quote comes from, comes from a, an adaptation of H.G. Wells' Empire of the Ants. And the original is, I for one welcome our new insect overlords.

**Chris Romeo (40:23)**: Yeah.

**Izar Tarandach (40:24)**: But the quote gained popularity after the 1994 episode of The Simpsons, Deep Space Homer. In which the news announcer Kent Brockman mistakenly assumes the Earth is about to be invaded by giant space ants. Fearing for his life, he announces his willingness to submit to the imagined invaders. So yeah, I for one—

**Matt Coles (40:45)**: So that Brockman effect.

**Izar Tarandach (40:47)**: It's the Brockman effect.

**Chris Romeo (40:48)**: There you go. Brockman effect.

**Izar Tarandach (40:50)**: Say it again.

**Chris Romeo (40:51)**: The Brock-man effect?

**Matt Coles (40:53)**: Brock-man effect. Brock-man effect.

**Chris Romeo (40:56)**: The Brockman effect. Gotcha.

**Izar Tarandach (40:58)**: Okay.

**Chris Romeo (40:58)**: I'm slow. It's early in the morning. The Brockman effect. All right. That's what we call it from now on. Patent pending trademark, whatever it is you have to do.

**Izar Tarandach (41:07)**: We're going to write a paper about it.

**Chris Romeo (41:09)**: And, uh, someone says it. Well, I think we, we got somewhere. I don't know where it is, but we got somewhere off a cliff. Thanks for listening, tuning in. Go to YouTube, click subscribe, send an episode to a friend.

**Izar Tarandach (41:24)**: Beware of the robotic arm.

**Chris Romeo (41:27)**: Yeah. If you don't send a clip to your friend, the robotic arms are gonna come and get you so we can control everything.

**Izar Tarandach (41:34)**: It's the claw. It's the claw.

**Chris Romeo (41:36)**: Send, send this to somebody else who needs to hear a bunch of, I don't know what to say.

**Izar Tarandach (41:42)**: Apocalyptical lawyer manager sci-fi.

**Chris Romeo (41:46)**: And then don't forget, don't forget the Brockman effect. Remember to thank your AI. Thanks for listening. If you enjoyed today's conversation, subscribe wherever you get your podcasts, leave us a review, and share the episode with someone who would find it valuable. We'll see you next time.
