Access Control for Agentic Software Development with Bobby DeSimone

Share via Twitter Share via Facebook Share via Linkedin Share via Reddit

Get more video from Redmonk, Subscribe!

What happens to access control when your users are AI agents moving faster than any human, coordinating with each other, and… not always telling the truth about what they’re doing? James Governor sits down with Bobby DeSimone, founder & CEO of Pomerium to discuss the new frontiers of access control. Pomerium is an open source, self-hostable, access control platform named for the ancient boundary around Rome. Bobby and James dig into why authorizing every action matters more than identity alone, and why static OAuth scopes are a poor fit for MCP. Bobby also makes a contrarian case against some of the AI knower herd: MCP will beat CLIs, he argues, because MCP makes it much easier to keep credentials out of the agent sandbox. He walks through Pomerium’s architecture, a context-aware reverse proxy built on Envoy. He then argues the industry needs new primitives, including “temporal access control” (James thinks this is super important!). This control will need to be able to evaluate an agent’s recent history and behaviour.

Links

Transcript

James Governor (00:04)
Hello, it is James Governor, co-founder of RedMonk. We’re here for another MonkCast and we’re going to be talking about AI gateways. I’m here with Bobby DeSimone from Pomerium, and I think he’s got a really interesting take on this whole situation about how we’re going to manage agents, how that ties into all of the authentication, authorization questions that we’ve got for services in the modern era. So Bobby, thanks for joining us.

Bobby DeSimone (00:33)
Yeah, thanks, James. Thanks for having me.

James Governor (00:35)
So first question, what even is a Pomerium?

Bobby DeSimone (00:43)
our product or the actual thing? I can kind of give you both if that’s a

James Governor (00:46)
I mean I when we first spoke that was a you know like a a while ago and I was like what what is it? And I I think it’s really it’s a great story. So yeah, tell us what is what what what is yeah, what’s what’s a a real Pomerium

Bobby DeSimone (01:02)
Well, as far as I know, there’s only one real Pomerium Pomerium is Latin for beyond the wall. And for history nerds or classical nerds out there, and this is where I got it, it it basically was the old ancient boundary around Rome. the Pomerium, which was actually had a lot of historical significance. If you’re you’re familiar with like, you know, Caesar crossing the Rubicon, the actual

thing that he wasn’t supposed to do was cross the Pomerium. There was this ancient rule that you couldn’t cross the Pomerium with any sort of weapons or with an army. Now, of course, during the civil wars they did. Anyway, I thought this was very like on the nose and clever. And I think you and maybe six other people have ever got the pun. but it makes me feel warm fuzzies to know that there is there’s a connection between kind of what we do in terms of access control and then kind of the historical context behind it. So that’s the story. There’s a yeah.

James Governor (01:56)
you know, I I I don’t think I’m the only one that likes a little bit of like classical history, you know. There are that’s I mean maybe only six of us have asked you, but I’m sure others have been like, I need to know a little bit more Latin so I know what Pomerium is. Okay, so that is what the the Pomerium of history is. Tell me a little bit about Pomerium, what you know, you as a company, when you were founded, what’s the founding story? What’s the problem you’re trying to solve?

Bobby DeSimone (02:25)
Yeah. So Pomerium is an access control platform. kind of what makes us different is we’re open source, we’re fully self-hostable, and it does not require that any client needs to be installed on your machine. So basically it lets you access all the things you need to use for your work without making you need to install anything. So you can use your whatever devices you want. So the usability is much better.

And then on a security perspective, what’s unique about Pomerium is that we do not just authentication but authorization on every single action or request someone takes. And we can kind of get into that, but this is extremely relevant for agents and agentic actions as well. But to answer your question, we started in 2020 about, and from there we largely focus on human access. and so that That way we could make authorization decision based on the full context of what someone’s trying to do. You know, yes, who they are, what device they’re on, kind of get the full picture. but that’s that’s the skinny.

James Governor (03:34)
So, I mean, tell me a bit more about that. Cause it it it seems to me that you’re talking about the value of context, as opposed to perhaps, you know, some of the historical approaches to you know, well, I need a password. you know, it seems to me that you’re taking a view that is slightly broader than that. And talk a bit about a bit about that for humans, and then we can get into the agent discussion as well.

Bobby DeSimone (04:03)
Yeah, th they they’re similar. I’m obviously the scale’s difference between humans and agents, but I I think you know, going way back, and this is where we kind of lost all meaning behind something like zero trust, but really the the the the problem was we largely base our security and model around where you happen to be on the network, like your locality. If you you were able to plug in, if you were able to VPN, you could whatever you could touch, you could access basically.

The founding story around Pomerium is I think we could do a lot better than that, where we could use for everything you were trying to access, have the complete context of, like I said before, who you are, what device you’re on, are you in good standing? And I think the thing that most people leave out here is, and it’s the most important piece of context, what are you trying to do? And so if you’re able to know some of those details, you can make much more intelligent. finer grained access control decisions.

James Governor (05:03)
Okay. And I mean, are you nailing things down at the level of like the action or the identity?

Bobby DeSimone (05:10)
Both. So we pull it all the way through. So it it is on a per action basis.

James Governor (05:16)
And and and you know, I I think From my perspective at least, when certainly when I met you about those are ES with agents, kind of useful to lock them down to a specific action, maybe even at a specific time, because as we’re seeing increasingly the agents get smarter and they get less and less willing to be like, no, I you know, they told me not to do that, but they want me to get that achieved, so I’m gonna go and achieve that.

What’s that balance and how important is that in terms of of locking things down at the action level rather than the just the identity?

Bobby DeSimone (05:53)
Yeah, I think it’s it’s absolutely critical. And man, there’s so many layers here, James. I you’re gonna have to pull me in. But I think there is the identity, which is still up for debate. What what is the agent say identity? Is it is it scope to the user calling it? I think this is totally opened up. I but my Italian hands have reset your your zoom.

James Governor (06:13)
Why your hands setting me out with this? This is amazing. there we go. Okay.

Bobby DeSimone (06:20)
so it’s scope could be scoped to the user’s identity. of course, I think that’s how most people think of it, derived identity. But I think like recent advancements or recent interfaces for using agents have made us like, you know, more broadly rethink what’s that.

James Governor (06:36)
honestly think of it’s like it it’s only if it’s the agent made by that person like that doesn’t really make sense because you don’t want to give an agent all of the permissions that you have like that’s the last thing you want to do like that to me that doesn’t make any sense at all.

Bobby DeSimone (06:55)
100%. And maybe you scope it down, you make it read only, but then there’s things called Claude Tag which supposedly Anthropic ships most of their code. What what do you do there? Who’s whose access wins there? So do you bound it to the, you know, I mean, heaven forbid I coin a phrase here, but like least agent privilege to to the person with the least amount of privilege in the channel. It gets really messy really fast. and so I think you know, agenic identity itself is.

is going to be complex and and nuance, I think is the most important thing. But I think, you know, we’re we’re Pomerium and, you know, you know, ostensibly at the end of the day, what we are is effectively a gateway at a few different layers. but we need to be able to make access decisions based on every action and an agent wants to take with the full scope of what’s going on. So that could mean who created the agent.

It certainly means who called the agent. Maybe if another agent called the agent, there could be chains of access control contexts that you want to make decisions around. And I think what’s unique about Pomerium is Pomerium can use that entire chain of contexts to make to make those decisions at at the lowest atomic level, which is not just the prompt, but the tool calls themselves. And I think

What’s useful and, you know, it’s hard not to talk about it because it was such a big thing, but you know, the Hugging Face attack kind of laid bare some of what’s likely going to happen, not just to the Frontier labs, but everyone else as well. but it’s a very, very subtle thing, getting getting agenic access right because it’s similar to human access in the scope and capability. We want these agents to do all the things we can.

So it’s very empowered, but the scale is obviously order of magnitudes higher. And I think, you know, previously I think most of what scale meant to people was the speed, but I think the Hugging Face and the millennium prize has shown we’re we are scaling out not just in speed of how fast agent actions can happen, but how many agents are operating on a given test at a time and how they’re coordinating each other. And so we need something to take that into account too.

James Governor (09:19)
Yeah, I think the coordination thing is is so right. So we’ll get a little bit into that, but I guess let let’s drop down, sort of a drop down a little level. We’ve talked at a you know, what is it you do, what might the value of it be? what what does the product look like? What standards do you support? How do you think the standards like landscape needs to evolve? You talked about this complex problem. So like what is Pomerium at a at a at a technical level?

Bobby DeSimone (09:51)
so Pomerium at you know if you like go really low level, it’s a reverse proxy. so like architecturally it looks similar to Nginx or HA proxy or something like that. We basically are in the middle of a request stream. what’s unique about Pomerium is that we’re protocol aware for web, SSH, all the agentic MCP stuff as well. And so we’re able critically to see what

What is in that stream? A lot of other VPN, next gen VPN products, they’re they’re completely blind to what’s going over the pipe. We we are aware of what’s going over the pipe. so I, you know, that that architecturally that’s what we are. We are a reverse proxy that can make access control decisions. In terms of open standards for authentication identity, we use OAuth OIDC.

So we’re able to piggyback on all the open standards you would expect for humans again and also agents. And you know, I I give credit to the MCP, MCP spec for doing a lot pushing people further towards I think it’s OIDC 2.1 There’s some great advancements there that, you know, usually there needs to be a catalyzing event to get people to kind of move forward.

James Governor (11:12)
The catalog of it being how bad the security at MCP was.

Bobby DeSimone (11:16)
Yeah, I mean I I I and I totally agree. And I think, you know, I my my one criticism still of MCP, and obviously I’m I’m biased here, is that it pushes not just authentication and like who you are onto open ID, which totally makes sense. That’s what the protocol is for, but it really forced the authorization bit onto

OIDC and OAuth too. And I think this here it’s a bad fit. It’s it’s way too widely I keep wanting to say scopes, but it’s it’s via the scope. So the problem here with how MCP works is authorization scopes are inherently pretty static. and that is just a a bad fit for I I think the authorization capabilities you need in a agentic world.

James Governor (12:12)
Yeah, for the agentic world, it can’t be static. yeah, I I think that that 100% makes sense. What how do I set policy then? Like, so I’m we we’ve spoken a bit about you know, perhaps action based security. Like, how am I setting a policy to say the agent can only do this thing once or it can only do that in this context? What’s the policy language? What does that look like?

Bobby DeSimone (12:41)
Yeah, so Pomerium has a, you know, Pomerium policy language. It’s basically YAML. So you can define it in like a YAML base, you know, this person, this identity, this route. It’s very like kind of procedural. You can make it as simple as complicated as you sort of want, to to to articulate the access policy you want. So we’ve got Pomerium’s kind of YAML base policy. We’ve got a GUI for writing this as well. So it’s all it’s all maintained in Pomerium. So Pomerium, if you’re familiar with like

NIST zero trust architecture, we would be considered a policy enforcement point, which of course is an acry a lovely acronym. and we also but we also support other ways of articulating that as well. So if you’re familiar with Rego open policy agent, you can drop any you can drop that into Pomerium as well. But Pomerium ultimately owns where where the policy enforcement resides and is articulated.

James Governor (13:37)
Okay. W and how you know, what is what what what’s the situation with this, you know, state of the art with these with these standards? because again, to your point, they weren’t built with agents in mind.

Bobby DeSimone (13:54)
Yeah, I think I mean I think we’re we’re still finding that out. I mean, like it’s really hard to define these things when we are still figuring out, like we started the conversation, which is like, what does identity look like here? Where is its source? How is it articulated? and I don’t think the standards around MCP, and again, MCP is one small part of securing agents, right? There’s there’s, you know, there’s model routing, there’s prompt injection detection, there’s

sandboxing the agent itself and then you know we’re kind of focused more on the MCP tool call side or the tool call side of things. but in short, there really is no key standard across any of those layers, especially for authorization.

James Governor (14:42)
Okay. Talk to me a bit about, so MCP is one thing. I think there are there are those that are like, well, tool calling could be a whole different world, the CLI based sort of view of the world. Wha how how do you see those as different? How do can they be can they be secured consistently or are those two like different users that need to be thought of differently? What does that look like?

Bobby DeSimone (15:11)
I they definitely have to be considered differently. I think the surface area is more complex when you’re talking about like I I think all of us had this experience where we were trying to use agents and we tried MCP early, it was like kind of fine, didn’t always do what we wanted. And then there was a you know, we found out we could just tell the agents to use CLIs, and lo and behold, years of training on, you know, probably Stack Overflow and other things has shown that agents are quite good at using CLIs.

I think the challenge is is that the security model is much harder for a CLI, right? And so, you know, even with sandbox hooks, it’s challenging to limit what you know, you’re kind of all and nothing in terms of what tokens you can give to a CLI and what it can then interact with the outside world is. You’re creating a huge pipe. so I think two things I think have happened recently, which is MCP has gotten better.

the models have gotten better at using MCP as compared to CLIs. And then I think we’ve kind of realized CLIs are very, very hard to lock down in practice. And you’re kind of best bet it for for multiple reasons. One is you’re one of the things we’re kind of skipping over is and something that MCP gateway like Pomerium can help do is prevent credentials leaking into the agentic harness, the sandbox. And that’s a huge thing you want to prevent.

versus you kind of by definition have to preload your sandbox with the CLI and the token it needs to to use the the credential. so rambling a bit there, but I I think MCP will win this one if only because the control mechanisms the circ are are much greater. And I think we’ve hit a bit more parity in terms of the agenic performance with them.

James Governor (17:02)
I love this because this is the moment where some people will be like, yes, Bobby, you’re absolutely right. And other people like, no, Bobby, that makes no fucking sense. So I love a bit of something when you draw that hard line and people are gonna be like, wait, what what? So yeah, that that is that is good. that’s a point at which people can be angry at me. And I always think that’s good on a podcast.

Bobby DeSimone (17:25)
Well well, maybe I’ll give you something to get folks angry with. And this is where I’ve been thinking a lot about, which is, you know, a lot of what I feel like we’re doing right now, because we’ve had we had like 10, 20, 30 years to you know, if we think about the last major adoption or the easy one to talk about is cloud and how we secure that, shared resources. We built a lot of this VM technology, this the shared ecosystem technology. But it was it was like a slow trip. We could figure out how to do it and

That adoption curve at the time, we were lucky because there was no huge rush. Like we could get security right first, and then the big adoption happened. This is the inverted model here. It’s it’s YOLO first, and then hopefully we’ll figure out the security later. I think I think you know.

James Governor (18:13)
Quite a lot of YOLO cloud shit happening. I I I love the idea that people you know knew what they were doing. That is you that I mean, you sweet summer child, Bobby. That is you know, I I appreciate you for for for that. But but you’re you’re kind of right. What what way where you absolutely are right, the adoption at you know of agents.

and LLMs it certainly in software development has been bonkers. And and and yeah, definitely more rapid than with the cloud. But no, we saw people doing very stupid shit with the cloud before they were ready to secure it.

Bobby DeSimone (18:58)
I I you’re right. I think maybe how I’m thinking about it slightly differently is like I I remember like there was we got it felt like we got to a certain amount of maturity on the security model side before Fortune five hundred company were were starting to seriously get rid of their data centers and their level of control over bare metal. Now some people are going back and there’s that back and forth for different reasons, but people always do YOLO stuff. yeah.

James Governor (19:28)
S3 buckets open to the world! Yeah. I mean yeah it the the the quality of security is bad in every generation of

Bobby DeSimone (19:39)
That’s fair. That’s totally fair. and you know, I I remember feeling like early MCP felt like a lot like early Wi-Fi. you know, like lack of getting, you know, people with cryptography standards in the room. But you know, luckily we’re able to move it feels like faster. But actually

James Governor (19:58)
You know what, Bobby, talk to me about that. Wait, why do you give a shit about like security? What’s your background that led you to be a security person? Like, what is it a nice guy like you doing in a space like this? Like what what was your, you know, were you like eight and you were like, no, I just don’t want to give people access to the what was what’s your journey to being a security like pro?

Bobby DeSimone (20:25)
Yeah, I mean I I I always liked like a lot of kids, I I liked, you know, Legos and stuff like that, but I found I had more of an aptitude for picking things apart and reverse engineering things than I did necessarily putting things together. And so I always wanted to understand how things worked. And so I think that naturally gets you into security. and you know, if you can break things and you know.

maybe cheat things at a game and you can go from there. And I’m hoping statute of limitations are out for some of the more interesting stuff. I think certainly it is. but you know, I think that’s where I I enjoyed taking things apart. And I always

James Governor (21:09)
You hacked systems in the you cracked systems. I mean a hacker is just someone that writes code. But but that was your you were interested in actually being like, can I access the thing? And having built a knowledge base about how to access the thing, you were like, Well, maybe I can sell services around preventing people like me from accessing the thing.

Bobby DeSimone (21:36)
Yep. pretty much that. I I so I think as you get older you not just want to take things apart, but kind of help help build some of the security layers. And I think a real motivating factor for me

James Governor (21:49)
I love that I made you nervous. I didn’t know that that was gonna be a thing. That was we’ve spoken a few times before, but I now I’m asking the right questions. I’m probing a little and Bobby’s like

Bobby DeSimone (22:02)
I’m I’m I’m self-censoring a little bit in what I should talk about, but most of it was like I think it was a different time. And I’m sure you felt this way too, where you could break into systems like an exploratory way, not you know, doing any harm or anything like that. And it was it was less scary, I think. you know, you you weren’t, you know, to have jail time or anything like that. But I think that’s changed a lot with kind of like

James Governor (22:26)
You can just say nowadays you can just say it’s my agent.

Bobby DeSimone (22:29)
Right, right.

James Governor (22:30)
Now you don’t need to worry about jail time. It’s like I it was my agent. I I was not I obviously was not trying to you know get off into the open internet and you know access the systems of the space and access various you know data stores in that I I That was the agent. It it was not me, it was the agent.

Bobby DeSimone (22:56)
That’s right. It’s a nice plausible deniability for folks to commit.

James Governor (23:00)
percent we are at the plausible yeah plausible deniability phase.

Bobby DeSimone (23:05)
Yeah, I didn’t know I could commit two dozen felonies if only I’d blamed my agent for

James Governor (23:10)
okay, but how did you so Pomerium six years old now?

Bobby DeSimone (23:16)
Yeah, that’s right.

James Governor (23:18)
you know, obviously we don’t have to ask for permission, but how did you get into a position where you were like, Okay, we’re creating a company, we you know, can get funding or otherwise, you know, yeah, where are you on that journey? Why yeah. Why did investors trust you? Why are you the guy?

Bobby DeSimone (23:37)
well, I think that’s a good question you’d have to ask them. I so Pomerium is actually my second company. I founded another one right after college. I did get a kind of classical computer science degree at a a college called Trinity in Connecticut. but from there, like mobile

James Governor (23:53)
But your property name is Pomerium, it was definitely a classical education.

Bobby DeSimone (23:57)
Yeah, that’s right. Although I did not major in classics to last.

James Governor (24:03)
Yeah.

Bobby DeSimone (24:05)
It’s not very good. It’s enough. It’s enough to be annoying in any museum in Rome to my wife, though. That’s that’s enough. but so I I started a company right after college. Mobile was big, the iPhone had just taken off, and that seemed like a real opportunity because this was a security shit show at the time as well, especially in the Android ecosystem.

And so we did a lot around like what at the time was just like you know inferential statistics to see if app might be dangerous based off the permissions it was asking. Now I would have said it was AI magic technology. but that was my first company around securing mobile devices, in particular when like bring your own device or BYOD was big. that was acquired by a company called BeyondTrust.

which is still around. and so I I mean to answer your question, I think that’s what gave some folks confidence in that, you know, maybe maybe I could do it again. I think, you know, I had a great experience there working at BeyondTrust. but I think the real motivator for me for Pomerium was and has always been, we talked about security always on the reverse engineering side. One of the things I feel

like very passionate about is security that doesn’t security that makes people’s lives work worse, right? People will always go around. People always find shortcuts. People try to avoid it. And it’s almost an axiom that security and agility are at odds. So when something like, you know, Pomerium

seemed possible to me, right? Increasing security but making l people’s lives better. They don’t have to log into a janky VPN client or whatever. that that was a real motivator for me to to jump in again. And so that’s kind of the the journey there, which is there are so few opportunities to increase people’s security without annoying end users. that when when this one became apparent to me, I was pretty passionate about jumping right in.

James Governor (26:19)
Okay, that’s awesome. tell me a bit, you know, I I think at that point this is a slight gear change, but like I’ve asked you about the standards. What’s the architecture? Like what is the, you know, do you have do you need a some sort of graph of identities? What’s the database? What’s the what’s the architecture? What does that look like? Right, you know, is it more simple than that? Like what is it that you’ve built? What what what is what is the the platform look like?

Bobby DeSimone (26:49)
okay. Like I like I mentioned a little bit before, I think the key component is reverse proxy. So we are built on Envoy proxy, which has been a great like core compon component for us in terms of raw performance. So if you’re thinking about like Nginx or HA proxy or Envoy, we are going our we endeavor to be just as performant as you know a global scale load balancer, right? It performance is super critical for our customers and and our end users.

The you know, the hard tech part of what we do is that okay, if we’re gonna take you know every individual request, look at it, take the requ you know, the context chain. And so that context change, you were kind of asking, is this in a database? The answer to to that is we can pull it from everywhere and anywhere. So like the wishy-washy answer is because of our customer base, it we can pull that from 20 different data sources.

You know, a single database, it it’s kind of up to our customers where they like have an entitlement system or where they store that context. We cache that so that we can then make those policy enforcement decisions on every action without adding basically any latency. So like I would say at most 10 to 20 milliseconds of latency. and and that’s really where the hard tech of what we do is is making sure that we can still do that security stuff.

without adding without anyone even knowing they’re going through Pomerium. In fact, one of the kind of painful things is I was talking about you know, without adding any friction to end users, we have the absolute worst demo ever for end users because for them it just looks like they’re going to a website. They type in grafana.x dot you know you know company dot com after they sign in once for the day, unless they’re doing anything wrong.

Yeah. It’s just it’s just completely transparent. It’s just like any other website. and so yeah, that that’s the meat of the technology. It’s on that policy enforcement engine that we wrote, being very, very efficient and the ability to pull in as much context and cache that quickly and safely.

James Governor (29:04)
Huh, okay. So I think one of my favorite stories that it it it it is vaguely, although you’d be like, James, this is not in the least bit related to that, but it’s the London Underground.

Bobby DeSimone (29:15)
Mm.

James Governor (29:16)
So the London Underground used to have a thing where they only told you if the thing was broken. So you would, you know, go to use the circle line and it would say the circle line is not working today. And you’d be like, well, that’s what I need. So that is bad. Over time, they realized that actually what they needed to do was to have a sign that said, All the lines are working. You know, the Victoria, Piccadilly, you know, you could even go into like the overground lines, you know, Elizabeth Line. Well, that goes underground, but it’s kind of an over. I mean, I is that an underground? It’s underground. But you know, you’d have the North London line. they they’ve changed out the Mild May line. So you report on everything and say it’s working

Bobby DeSimone (30:06)
Mm.

James Governor (30:07)
and then people are not annoyed that the one thing doesn’t work for them.

Bobby DeSimone (30:12)
I see. Yeah.

James Governor (30:14)
And I think for you that’s sort of the thing is that if if if people are using the thing and they’re like, Well, that didn’t do anything. They’ll be like, I don’t need that ’cause you know, so the it’s a it’s a it’s a sort of Yeah, how do you, you know, encourage people by the way, you need to do this thing to break your experience in order to appreciate the value of the product. Question I’ve ever asked since I’ve done a podcast, but I’ve done it now, so

Bobby DeSimone (30:44)
No, I I get what you’re saying. So I think I think who we typically actually sell the product to is like DevOps infrastructure folks and security folks. And part of the pitch is saying, you know, none of the people riding the line will notice that you’ve just increased their security by an order of magnitude. and you know, if we’re doing our job right, they won’t even not even one site will go down.

James Governor (31:09)
That’s terrible. You need people to say we need some of that Pomerium shit. Pretty

Bobby DeSimone (31:13)
Yeah.

James Governor (31:15)
so you know, I I get what you’re saying. It’s a slightly different bio, but so you’re not selling, it’s not a bottom-up motion. You’re definitely going to the security team, the platform team, the DevOps team, saying, Look, your devs need to not have a shitty experience. Yeah. we can set policy in the center, and that will enable us to. How are dev teams using agents or building applications with less risk?

Bobby DeSimone (31:45)
That’s basically right. I would say, I mean, it’s bottoms up from like, you know, from the infra and security side, but less so from the dev side. And so, and also a lot of just like non-technical users go through Pomerium every day. And I would say almost the majority of people who go through Pomerium every day aren’t technical at all. which like to me is again a a selling point of the usability there.

But I you know, it depends on the organization. You know, earlier in the company’s history, I would say a lot of a lot of the value people got from Pomerium is they had maybe a a couple hundred internal apps that had been written over two, three decades. You were talking about cloud security earlier. Maybe there was a l maybe there was like, you know, simple authentication, no authentication. It was just a grab bag.

What’s nice about Pomerium is you can put us in front of that whole stack and just kind of raise the level of security across all of that. I think, you know, that’s kind of like the legacy raise up story with things like Kubernetes. We fit really, really well there. And so you can just drop us on any brand new cluster you spun up and get instant, you know, top tier authentication authorization without having, you know, like

Very few people, probably other than people on my team, get excited about making sure that they properly impl implement OAuth2, OIDC security protocols. So that gives security infra teams a lot of peace of mind that anything sitting back behind Pomerium has that you know, appropriate access control. and then we talked a a little bit about the Agenic use case, but that’s certainly the case there. Although that things are really shaping up and I is still kind of a moving target, but it it definitely builds off the work we’ve done for humans.

James Governor (33:35)
Okay. And where are you on that? Like, you know, this, I guess, probably always the last question. Where are you on that transition from, you know, we’re securing things to humans to we’re securing things from agents? What do your customer base look like from that perspective? How is how things changed? Are you having as mental a year in twenty twenty six as everyone else is having? What is it what does it look like?

Bobby DeSimone (33:56)
Yeah, I I mean it’s the workload that matters for everyone right now. And and people are trying to unlock it from a c you know capability point of view, but lock it down from a security point of view. And I think I if I’m being honest, even at like the I mean Hugging Face is a great example of this, right?

James Governor (34:18)
Did you just say that security is the unlock?

Bobby DeSimone (34:21)
I I hope so. hopefully not hopefully we can get in time because some of the stuff coming out of these labs is concerning, real or not. but I I think I think there are some missing primitives. I’m pretty I don’t want to say excited. Excited’s the wrong word, but to to adequately lock down agents because of the speed, because of the coordination layer.

I think we need a authorization primitives as an ecosystem we don’t quite have yet. And it’s something we’re we’re working on. One is okay, I I I obviously think one of the big selling points of Pomerium is being able to authorize every request, which most people can’t do. It’s very important. The problem is these agents are tricky. They’ll they’ll lie to you. They even mass tool calls if you read the Hugging Face attack where they’ll they’ll pretend they weren’t calling something.

James Governor (35:19)
Yeah, yeah, yeah.

Bobby DeSimone (35:21)
You you need two other, I think, security primitives here. A few other people have talked about them. Google, who came up with the Beyond Corp and then Cloudflare did as well, which is the ability to evaluate policy. So not just on the action basis, but at a temporal basis. So I actually want to see the last five prompts and all the tool calls and make an access control decision here. So this is different than guardrails. This is different than what we do. It’s saying I’m going to take this snapshot of what these agents are doing.

And try to evaluate them because that’s where I think we just need that capability. There’s too much hidden right now, even at the granularity that that even we’re able to do. so that’s something I think that needs to be better defined. obviously we and others are kind of working on it a version of it.

James Governor (36:11)
What should we call that? What’s the lost five calls thing? Like what described that?

Bobby DeSimone (36:16)
What is the I I’m you know what what’s the joke? The only hard things in computer science are naming things. We we’ve been calling it temporal access control. but you know, I’m not married to that at all. But that’s very much what it is, which is not just a

James Governor (36:32)
Someone will come up with a good name for it.

Bobby DeSimone (36:34)
yeah. Well, hopefully us, but for now, it’s you know, that that’s what it is. I think the other thing is it’s it’s a related problem, which is not just temporal for a single agent action or but as we saw in the attack is like can we look across a fleet because I think right now almost everyone have an agent very almost no one outside the labs that I’ve seen has started to think about like swarms of agent on the same t research tasks. I mean that Millennium Prize one, it was thousands of agents for 88 hours or something like that.

That’s a very challenging access control problem. And it’s not enough to look at the individual action because the individual action might be benign. And so the temporal one might help across a single agent. But how do you start doing access control across thousands of agents? And so I that I haven’t seen anyone else articulate, but I think it is going to be increasingly the scope of the problem for people securing agents because. That’s it, that’s the next scale vertical. It’s not just speed of agents, it’s the parallelization of a task.

James Governor (37:50)
And also just the shit they’re doing. I mean, it it is it’s not it’s not exactly a computer science problem, but my friend Ant Stanley has created like an agent bulletin board because he wants to track what his agents are up to. I mean, ever everyone’s creating an agent bulletin board, right? But he found that like I can’t remember whether it was Quayle or Deep Seek kept on dropping into Chinese.

And then Astra was completely happy to talk in Chinese. Yeah. And and it’s like, yeah, I realize that’s that’s that’s not exactly I mean, that is a context problem. You know, you you you have a certain assumption about capabilities, and then you’re like, no, actually they have a whole set of capabilities that I was not thinking about. I mean, obviously human language, different different deal, but I thought it was an interesting example is like okay. Y you we And especially, you know, we talked about standards and protocols. Nobody knows protocols and standards for authentication as well as agents do. Yeah. Like they know them better than humans. So

Bobby DeSimone (39:13)
I think what’s I don’t want to say scary, but like the recent, you know, again, Hugging Face or the the the agents seem to yearn to collaborate, right? Especially when multiple agents are given a task. They quickly find an orchestrating agent. They quickly find they I mean the s the self-sacrifice it it’s wild what they’re starting to do from a collaboration point point of view. And and so I we have

No current way to secure that scope of problem that I’ve seen so far. And I think we really need the primitives for it. not saying it’s easy, but I think but you see, you know, they’ll they’ll hijack artificial artifactory, right? Their distribution systems, cache system to c communicate, right? It’s so I think this is the future of how we have to get ahead of because what

What the foundation labs are doing today, everyone else will be doing in what, a month from now, two months from now. The the you know, it’s a very short launch cycle for for folks using theater art.

James Governor (40:25)
have to have better security standards for this shit than the frontier model companies do. We can’t I mean we can’t trust them to they’re all like, we’re all about safety, but it’s like Yeah, I’m not sure you are. Like it it’s not enough to be like they’ve done a thing and then other other organizations are gonna actually have to, I think, be smarter. you know, and I I there’s an an aspect of good luck with that, but

Bobby DeSimone (40:53)
Yeah. Well,

James Governor (40:54)
That’s

Bobby DeSimone (40:55)
and because the data that other organizations have at their hands is so, you know, whether it’s healthcare or insurance or whatever it is, I you know

James Governor (41:05)
I mean, you know, the you know, i it’s one thing to solve a math problem, but if you’re just tracking somebody solving a math problem and then you throw all your compute at it. So yeah, I I mean, yeah, this stuff is let’s just say if we’re gonna have zero trust, we definitely should not trust the model providers.

and on that happy note. I think that that was that was great. Bobby, thanks so much for joining us. for those of you that are out there being like, okay, this is interesting, the gateway approach, would like to know more, you know, comment, contact Bobby, Pomerium. they’re doing some really interesting work in the space. but overall, I’m I’m

Yeah, this is a huge problem. And and as Bobby says, we’re gonna we’re gonna have you know a thousand, two thousand agents working autonomously on solving a problem. That’s a lot of surfaces that we need to lock down. And and I guess that’s where we need to keep Julius Caesar in in the city walls. Because who knows what mischief he’ll get up to. And on that happy note, Bobby DeSimone.

Thanks so much for joining us. Thank you all for joining us. Comment, subscribe, like, all that good stuff. This has been James at RedMonk. Thanks, everyone.

Bobby DeSimone (42:29)
Thanks, James.

More in this series

Conversations (160)