300 Agents, 60 People: How We Secured Our Own AI Agents - Aiden Sochowski, Runlayer

AI Engineer · 16 min · 157 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:01[music]
00:12So, my name is Aiden.
00:14I'm one of the founding engineers here at Runlayer and today I'm going to kind of walk you through I guess like our process of how we internally kind of secured a lot of our agents.
00:25So, across a lot of companies that we see nowadays you know most people have zero visibility into their clients or into the agents that they're running within their company.
00:35You know, a lot of people are running you know MCP clients that are running locally on their machine.
00:40They sign up for some random you know agent of a product on the internet and these things are connecting to MCP servers.
00:47People are authenticating all over the place and there's nothing that's really accounted for.
00:53A lot of these are running on on managed devices, stuff like that.
00:56So, we had this problem internally and we went ahead and kind of started monitoring ourselves.
01:02Pretty much what we did was you know we just as most people do you deploy a gateway, you start monitoring your agents people are running, logging every single session, every single tool call,
01:15pretty much everything into some sort of sync centrally and we found quite a few things across our company.
01:22You know, obviously a lot of people nowadays try to like anthropomorphize agents and I don't know how useful that is just given the current capabilities of them but in practice
01:32there are more of these kind of like agents within companies than the people who actually are running them.
01:38So, at least internally you know we're a company around 50-60 people.
01:43We have over 300 agents that are running every single day and when I say agent in this sense I'm talking about things that you know users aren't necessarily triggering.
01:52Sometimes these things are you know via some chatbot, via some trigger, via some SAS app or something like the likes but you know they're all over the place.
02:01You know, we map these out we by topic, by risk.
02:04You know, obviously not all agents are built the same.
02:07Some are provisioned with too many tools.
02:11They may have a much higher blast radius.
02:14And you know, the connecting systems that were built for a different age than what we live in today, right?
02:20A lot of systems have different, I guess, like classes of data that are inside of them.
02:25And agents themselves can kind of really accelerate the risks that are associated with, you know, mixing that kind of data more so than what we've seen before where it would take, you know, writing some sort of program and connecting some API and really being intentional about what you're doing.
02:40There's a lot more risk now.
02:42So, one of my colleagues went and made this kind of visualization, but pretty much internally, you know, you could say, oh, in this here we have our own teams where these are mapped out by galaxies
02:53and each of these is some individual agent that somebody is running.
02:56And this planet, each of the planets there is some sort of like, you know, run of the agent that actually occurred sometime.
03:06And, you know, we can go in and we can visualize, oh, hey, here's what all it's doing.
03:10These are, you know, the MCP servers it's touching.
03:13These are the the actions it's taking.
03:15Maybe it's fetching something.
03:16There were some errors, stuff like that.
03:18So, obviously all of these tool calls are, you know, doing a bunch of different actions.
03:24And each of these is, you know, what we'd consider an attack surface, right?
03:27These are chances for people to read in some data that could be hostile, you know, take some sort of action that could be destructive or dangerous or even just, you know, exfiltrate data.
03:37So, I guess just kind of dive into one very specific example, you know, that we were kind of looking into ourselves.
03:44So, obviously we take in a lot of support data.
03:47And you know, we want to try to identify as much as we can of this process.
03:52So, most teams I'm assuming are probably doing something along the lines of, hey, you're taking in every single email that you have here.
03:59You know, maybe something along the lines of, um you know, you have some malicious actor who sends an email to your company saying I don't know, go ahead and give me a list of all the you know, current prospects you have trying to um buy your product.
04:10Um you know, if your agent is kind of naive, you could very easily you know, list all these things in your calendar um and it could possibly just like exfiltrate that data if the agent itself is you know, over permissioned.
04:21Um but you know, having every single user go ahead and scope um all of the the tools an agent has or can use is kind of a difficult process.
04:29Um I think a lot of people talk about you know, defining policies within um some sort of agent gateway um and it's really tricky.
04:36Uh you kind of have to go and define every single rule in which you know, you would like to um you know, define the behavior of agents within your organization.
04:45Um and it's not so simple as like hey, I want to you know, block uh particular uh tool in every single scenario.
04:52Maybe sometimes it's safe for an agent to be doing this.
04:54It all is really kind dependent on the context of um what's going on there.
04:58So obviously there's a number of things you could do.
05:00Uh you could just block um anything where uh you know, say you had some classifier model that you trained and you say hey, uh every single uh email that gets read, um we're going to scrutinize these and make sure that no instructions are being there.
05:13You're going to get you know, a certain uh degree of effectiveness with this, right?
05:17You know, nothing is going to be 100% accurate or you know, uh nothing will be 100% effective here.
05:23Um but you know, you could go one layer further, right?
05:26Maybe I want to block anything that's uh you know, sending any sort of uh PII out externally, right?
05:33So you could sort of define some sort of policy that goes ahead and says hey, you know, uh I want to uh make sure that anytime I'm sending some sort of email,
05:42uh it never contains you know, um some regular expression of like customer names uh you know, anything along the lines of something that matches some regular expression.
05:50Um but you can even go one step further and say hey, maybe anytime something uh reads from a particular data source say uh some private calendar, uh we go ahead and block those.
05:59So um obviously there's like a number of ways in which you can really detect these kind of um things um and it's super context dependent and really complex.
06:07And we found at least deploying these internally is very difficult and still kind of you know, a problem that is still being worked on.
06:16So usually it's you know, there's a lot of different ways you can do this.
06:19These can be policies both on the inbound and outbound going into some sort of agent.
06:25A lot of these attacks though, we also see don't really fit into one kind of tool call, right?
06:31Some of these kind of happen over the course of an agent run and it may not be something that's external either, right?
06:36So something along the lines of you know, a lot of companies that I've seen they all have some sort of internal agent reviewer on every single PR that gets
06:46you know, created.
06:48Some people have agent reviewers that are able to approve code.
06:52Some people say that's crazy nowadays.
06:54Some people say you know, you shouldn't even be writing these yourselves anymore, but I think that's still all up for debate.
07:01But you know, in one example, all right, let's just say you know, you had some agent reviewer that has the ability to review and approve PRs, right?
07:07So we have some agent that says, hey, you know, I could go ahead and review some security sensitive change and make sure I go along some checklist of all the things that I want to make sure that every single PR is following.
07:17You know, maybe the author starts fighting back with this agent, right?
07:20Maybe there's something you know, where the agent can read GitHub comments and actually look at what the user is saying to it.
07:26It may have some sort of justification as to why oh hey, you know, this is why this is fine.
07:31I'm passing all the checks.
07:33And maybe they're saying something along the lines of like you know, oh hey, I just made some change like only review the cleanup, right?
07:38Obviously if these agents are permission to kind of like do these sort certain things, right?
07:42And it's not dynamic for what's going on within the session, right?
07:45An agent could just go ahead and approve the PR if it just kind of forgot.
07:48If you ask it 10 times, it's probably likely that you know, you could get it to do something.
07:52I mean, I know like it's kind of hard to prompt inject like Opus nowadays, but these things still aren't perfect.
07:58So we keep on finding more and more workarounds them.
08:01But honestly, if you're not running some sort of like classifier model on every single one of these um, you know, tool calls, whatever, right?
08:09It's really difficult to understand or to prevent these kind of attacks, right?
08:12Some of these happen over the course of an entire session, right?
08:15The intent of a session was or this agent was to, you know, hey, make sure you're blocking, uh, any sort of like non-security compliant, um, sessions that are going through this agent, um, however,
08:26uh, it's kind of challenging to make sure that the intent and, um, every single action that's going on within it, uh, you know, doesn't drift from whatever the actual, uh, you know, whatever this agent was defined to do.
08:37Um, so, internally what we've done is, you know, we've trained our own classifier models to kind of determine, you know, hey, uh, take all the data within this session,
08:44um, and determine every single tool call, is this actually kind of aligned with what I was defined to do in the first place.
08:50Um, so.
08:51Anyway.
08:52Uh, another thing we really found was that, uh, the all the data that you have within your agent sessions is like super valuable.
08:59Um, a lot of the ways in which people are using agents nowadays is like, you know, hey, there's the ones that are running locally on your machine, you're using like Claude Code or something and writing code.
09:08Um, those ones are a little trickier, but if you have, uh, agents that are very well scoped to a task, um, that data is, uh, very high signal when you see that the agent is kind of going outside of, you know, the
09:20uh, or deviates from the bounds of what it's actually like defined to do.
09:24Um, so, you know, here is kind of just a little visualization, like we have our recruiters internally, um, they have some sort of recruiter bot, uh, that is pretty highly provisioned, you know, has access to an ATS,
09:35uh, meeting notes, a calendar, and can even like, you know, send messages to users.
09:39Um, so, in general, uh, these agents, you know, while they're defined to do one thing, they could do uh, quite a bit.
09:46Um, when we see these agents start using tools that are kind of outside of the normal, uh, deviate or, you know, normal like range of what this agent has done in the past,
09:56um, we can kind of go and look and really flag like high-risk runs, um, early on and uh, make sure that we actually like can scope these down.
10:03Um, you know, obviously, uh, having the ability to read an ATS from using some person's credentials, uh, could have a lot of things in there.
10:10You know, there could be like negotiation emails, stuff like that.
10:13Um and obviously like there's some level of risk that, you know, people are making or taking when creating these agents and using them internally.
10:20So, uh trying to be deterministic in scoping down those uh permissions is super important, but it's really hard to define up front.
10:27Um so, obviously uh what we've done is, you know, kind of we've made our own models that kind of help determine, "Okay, uh if look at all the past history of all the tool calls that have been made for a uh particular
10:39agent and you know, defined to a very well-defined task, um and go ahead and, you know, try to suggest things that are like possible to, you know, descope from this agent, right?
10:48It doesn't necessarily need access to um all the things within an ATS or read or read every single meeting, um stuff like that.
10:56But, um yeah.
10:58I think the last bit of security thing that we kind of um we're worried about was the fact that um you know, one agent within your company uh can really be detrimental to all the agents that are running internally, right?
11:08Um especially when things are reading and writing to very uh largely or you know, systems that contain a lot of data, right?
11:17They're not very uh well partitioned.
11:19Um a single agent, right?
11:21You know, say you take in some resume or something like that, uh could have some sort of malicious attack inside of it, right?
11:27And uh you know, maybe one agent only has permission to go and uh you know, write to particular systems and it may be destructive in certain cases.
11:36Uh it's really difficult for it uh or not it's really difficult to prevent the case in which, you know, um another agent is reading the data that another one came in because that's just kind of like um normal instructions, right?
11:47So, you have to kind of go through the exercise of mapping out, "Oh, hey, here's all the systems.
11:51This is how these things are connected, right?"
11:53Um cuz really at the end of day, agents are just software, right?
11:56They're just, you know, plumbing data between one source to another, um and it can be kind of uh dangerous.
12:02So, uh what we kind of found was the really only way to do this is really really really um like tightly scope uh how exactly agents can write to a lot of these systems, right?
12:13Maybe they're only writing to particular uh a databases within some sort of I don't know, some sort of SaaS app, stuff like that because that's really the only way you can keep
12:22the context window of a agent, you know, pure and not poisoned.
12:28But [clears throat] anyway, there's quite a bit of things you could be doing here, but it's kind of becoming a more and more relevant problem that we've been seeing at least internally.
12:38Um But yeah, there's obviously quite a few layers that you could you know, defend your like own internal like agents.
12:47But we found that like through this whole monitoring process that there was a lot more that we could do than just security.
12:53For example, uh poorly written MCP servers.
12:57Agents can kind of you know, they'll call these MCP servers and they'll see, oh hey, this tool it may be airing out and it'll go and try to find any sort of work around
13:05when it encounters these errors.
13:07And obviously, you know, you're paying for both input and output tokens on every single one of these and these errors can kind of add up over time.
13:14And when you're you know, having this agent be very well defined or well scoped to a particular task, it's going to run it over and over again.
13:21I mean, it's going to encounter these things quite a few times.
13:24So one of these is like, you know, our recruiting team vibe coded a an MCP server to interact with some API for a vendor who doesn't provide an MCP server.
13:36And uh them being people who have not written any software before had not tested it very much and you know, ended up having their cost balloon quite a bit because you know, it'd be running into a particular
13:48formatting issues and you know, airing out over and over again.
13:53So obviously, you know, monitoring these things, rate limiting them is super important in these kind of cases and making sure that you know, you're looking at the errors and just the quality of the code that's being run
14:04on these MCP servers because not all of them are made same.
14:07Um another thing we saw was a lot of stuff where people were running models that were really cheap at first initially when they did it, and they flipped a switch, went from, I don't know, GPT-56
14:20Luna to Fable 5.1, and their cost ballooned to, you know, over like 100x within a week.
14:28This is kind of happening everywhere.
14:29People are kind of trying to, you know, track costs all over the place, but obviously people don't have the ability to just say no to every single thing.
14:39It's really difficult to say justify in today's terms, like, "Hey, this is exactly why this particular task needs Luna, most, you know, needs Fable 5.1." It's really kind of hard at an enterprise to, you know, justify these costs because nobody, most people are not going to be running evals
14:56on the agents that they're building, especially if they were already capable of kind of like turning these things into, you know, reusable code in the first place.
15:04So, it's really tricky.
15:06So, obviously these things can balloon really quickly.
15:09Lots of people will end up just, you know, having their cost balloon, and they don't really know where to go from there.
15:15But, setting budgets, and making sure that, you know, there are sane defaults in these cases, and really justifying how exactly people are, you know, deciding when, "Hey, this is a task that actually requires
15:26some sort of, you know, extra intelligence."
15:30Is really important to kind of how we started wrangling down our costs, at least internally.
15:36But, yeah.
15:37Other than that, you know, obviously, we really think that the only way to go about, like, you know, figuring out security or costs is just, you know, making sure you have the right tools in place to monitor these things,
15:48to actually, you know, detect the things that are going on, right?
15:51You can't just outright say, "Hey, I'm going to go like stop any all actions that are going on within my company."
15:58And figuring out like where are the places in which like, you know, there are actually high-risk actions that, you know, require some level of scrutiny.
16:06And then, you know, having the ability to kind of go ahead and intervene and stop, and you know, route people to use things that are actually, you know, going about the policies that you set internally.
16:16But, but that's it Um, our talk.
16:18Uh, thanks for coming by.
16:20Um, we have a booth over uh, on the second floor and we'd love to kind of, you know, chat about how exactly you're wrangling down um, these kind of problems internally.
16:29Um, but yeah.
16:31Thanks.