Protecting AI agents with Cloudflare One

Secure AI

About the session

AI agents are proliferating inside enterprise networks with static API keys and no identity checks. This session deconstructs how Cloudflare One treats agents as first-class identities at the edge. Explore machine-to-machine authentication via Access service tokens, short-lived mTLS, and ZTNA segmentation enforcing least privilege without network complexity. Also, learn how you can use Cloudflare Gateway to inspect agent-to-SaaS traffic in real time and DLP policies to prevent exfiltration. The outcome: a concrete architecture that replaces secret sprawl with policy-driven enforcement.
View transcript

0:02All righty. Hi, y 'all. My name is Ayush. I'm joined here with Todd. This is the Protecting AI Agents with Cloudflare One. It's going to be about 20 minutes. We do have a live demo and a little bit more hands -on stuff that we'll show you on how to do this. So strap in. If you have any questions towards the end, we'll be here afterwards as well. So please, you know, save them to the end because we have a little bit of material to get through. Before we I'd like to introduce ourselves. I'm a Principal Product Manager at Cloudflare. I deal with a variety of things, but specifically thinking a lot about AI agents and how we do behavioral analysis and things along that side for them. Yep, Todd Murray. I manage a group of solutions engineers who are out of the Americas. We're responsible for AI, Zero Trust, SASE, email security, things along those lines.

0:51So, perfect. What you're here for is we wanted to kind of start off with an example of what you may be running into within your organizations is you set up an agent that's supposed to read an internal record it's supposed to prepare a summary and send it to an external endpoint right a very common agent probably one that you may be running already internally or looking to actually deploy but what happens if the agent accidentally sends the hold record right it's supposed to only send a small part or summary what is going to happen if it accidentally sends the whole thing. We talk a lot with customers that are really worried about the fact that these agents are non-deterministic. It's really hard to know what their expected behavior is going to happen, right? And so how do you prevent this agent from going off the rails by providing it some of the guardrails?

1:38And this is kind of what we're going to be talking about in kind of the demo.

1:43So we see a lot of this when we talk to customers and how they're kind of deploying it themselves. For example, oftentimes they'll use a secret, they'll use a token, right? These tokens can... sometimes kind of outlive the tasks that they're used for, right? So you have an agent that has a token, it was only supposed to do it for a short amount of time, but now this token has continued to exist and maybe misused or the agent can continue to use it over and over again in different ways. You know, you have the network access becomes application trust. You have the full record leaves unchecked and then decision -like usable evidence. So kind of going through those things, what those actually mean is It's really hard to know, for example, when you're decrypting maybe like a website, right? You're trying to look at kind of traffic going through your network.

2:30You have to do, you know, all these body inspections or requires TLS decryption. But it can still, if you have an authorized read, it can still lead to kind of an unsafe kind of upload, right? So how do you prevent your agent from doing things that it's not supposed to be doing and to make sure that you've really scoped it very cleanly? to the task that it's supposed to be doing so we kind of think of it in two questions and kind of two controls here so what way what may my agent reach and what may my agent release right and so the way we split this up and we'll see we'll see more and actually the demo itself is on one thing on one side the first one what may my agent reach is really the authentication and authorization problem right what is it authorized to read what is it authorized to actually view what is the data access to, right? And the second one is what made the agent release, right?

3:16Do you want your agent to be able to send out PII information, right? Do you want them to be able to send out sensitive social security numbers that it may be reading from somewhere or PII information like health records, right? We see this often, right? And the way that we conceptually like to talk about with our customers is these are two different kind of problems, right, that you're trying to deal with when you're building your agents. And we'll kind of get into the products So we have our access and kind of gateway DLP product that helps those two different things. So we're actually going to do this demo. We're going to show you a live demo of this. The first one, excuse me, is a headless agent. And it's going to go through our Cloudflare Edge using our Cloudflare tunnels or Cloudflare D. And then we're actually going to hit a private application like a private API.

4:05Excuse me. And the second one is we're going to use our Cloudflare 1 client, which is our warp client. And that's actually going to proxy the traffic. through the Cloudflare Edge, using Gateway and our DLP. And this is the second question, right? How do you control what's going out there? What can we kind of analyze, whether that's the body, what the contents of the request is? And then we're going to go to a controlled external endpoint. So these are the two things we're going to do a quick demo on. I'm going to hand it over to Todd. We're going to walk through his setup and how these things actually work like in practice and how you can set it up yourself at home. So I'm going to show you the... demo, but first let me kind of describe the lab. I built a lot of these. This one in particular is Proxmox hosted inside of GCP so they can do layered virtualization, right?

4:53From there, effectively, it's designed to look like five or six different data centers. It's all interconnected with SASE. You can see a whole host of routers down at the bottom. All those are running our own Cloudflare WAN technologies for connectivity. Then on top of that, I've got some mcp services to manage the proxmox environment do a lot of the builds i've got some clients to actually reach out and and send data back and forth and then finally you've got the demo client there and then the one down here and those are the two that are kind of a deterministic state that i put in there in order to send data back and forth right so that's what we're going to focus in on um for fun and i'll show this at the end everything i built i built the AI. I've also put it in our internal GitLab, but I took it and I also published it out to GitHub if you wanted to basically duplicate this as a test bed.

5:41So you can set it up, it'll walk you through all the steps to setting up access and MTLS, perform all the security elements of it, put it in the live lab, you can test it, see the data, the requests, and then adopt that to production use, right? And we'll go through that. But here's kind of the page that we put together. I'll show you the actual logs, but what it's going to do is it's going to intercept those things. We're just going to look real quick at the different use cases, right? So the first is how do we establish identity for the client? In this case, we're mostly doing it around MTLS, right? So we're going to use certificates to validate that the machine should be able to have communication. We're using private infrastructure, right? So PKI is built into the demo.

6:27And there you can just kind of see there's the request. If we're not sending anything, we're just verifying the actual... issuance. All right. On the next, we're going to send a request and we're going to do this without authentication. Right. As you would expect, it's not going to get where it's supposed to be going. Right. We're going to block it. There's going to be no response whatsoever. All right. Now, tokens are great. They're a great approach. What they don't do is necessarily validate the origin server. Right. They validate that the request came from somebody that it probably should, but you still don't know whether or not it came from the right infrastructure. So if I separated my machine from a certain agent, for example, right? You've heard about Cloudflare OS. If Cloudflare OS was talking to an internal service, how do I separate it from me?

7:12This is how we would do that is on the TLS side, right? So token only, same thing, we send the request, we send the token, we know it's valid. We're still not gonna get a good response, right? Take the next step. Here we're gonna send both the token we're going to validate the certificate as part of the actual request all right so there's your get and now we're actually going to get back to 200 which means those two services are going to be able to talk correctly so we've verified the inbound direction we now know it's coming from the right service we have token for authorization but we have tls to verify the actual origin all right so next we're going to send data make sure that goes through great all right we can post the data out so we push stuff in right so this is our response going out to a service this could be an mcp it could be a sas service whatever whatever that tooling is but this is your outbound right and then the last step is let's look at the response side the problem here and this is why it's kind of important is because when in the lm world you don't necessarily know what that AI agent is doing. It's doing all its work. It's going to skim past.

8:25You're going to see some logs. You don't have any time to interrogate that. You certainly don't have time to look at the response and interrogate that to see if what is coming back to the AI from those tools is appropriate. Is it the right data? Is it too much data? Did I just cause an incident or a breach because I pulled, you know, instead of a first name, I just pulled somebody's entire health record. right it's really easy to make that mistake you might not find out until you're in the news because suddenly your data is showing up in a generative AI right that's what we're trying to avoid here so in this case what we're doing is we're taking the authorization now we're on the response coming back from the tool before it goes back out to AI and we're able to actually look what that data is and block that from a DLP perspective right no human interaction is involved in any of this

9:23right now on this if you do want to take it and use it feel free I just pushed this over today it should work I'm gonna go with should because I didn't test it but effectively what I did is I took this demo lab I put it up at github so that you could try it and use it up right

9:49On the log side, mine is light in there, but this is just to kind of show you're going to see the same stuff on the log side, right? It's going to come in the clefler axis. You can see the actual requests, whether or not they're forbidden or whether or not they're allowed. Here you've got my OKs, right? The actual JSON requests behind those payloads. So the proof is pretty easy to get to in that case. All right, like I said, I built this all. off of ai so it's easy to rebuild and then it is published out so if you didn't get the qr code it's just mellow techie it's kind of my handle is where all my sites and my demo domains look but you'll just find it right there okay back to the deck please yes black back to the deck

10:39cool so you know we had 20 minutes so we kind of blazed through that a little bit but really to go back to the core idea of this is that we just showed you five different outcomes, something that, you know, your agents will be doing maybe in production, maybe just an internal capacity. But the first one we showed was with no credential, obviously very simple to do, right? Access denied. But the next one, especially as you're hitting these private API endpoints, whether that's an internal server, or you're running that you want your AI agent to maybe get health stats for, it could be a plethora of things we've heard our customers talk about. You can see that just having a token, you can an access deny, the API denied it. But if you want to make sure with MTLS that you have both the token and the valid certificates, that they can have that conversation, you can do so there. And you saw that in kind of the third example. And then finally, a safe post, right, which is, hey, I'm querying my Salesforce, right, and I want to pull some information from it.

11:29But I want to make sure that, you know, my agent is not pulling sensitive information about that Salesforce record and then feeding it. to a generative AI that, or a generative LLM that now is being trained on that data, we showed in one way, you know, without any sort of sensitive information, it went through, but then in our ability to block that by inspecting the traffic, we were able to show that in the end that we were able to see, hey, this had some PII, some, you know, information that should not be leaving based on a policy and we were able to block it, right? So that was kind of the big TLDR is like, look, we gave you some examples, obviously you have Todd's. of uh test test uh github that you can do and spin it up and play around with yourself when using tunnels using access using gateway and using dlp and all these different building blocks to kind of play around with it yourself on on different uh use cases that you have um and so kind of how a little bit deeper uh and we're going to kind of go through a little bit fast as well because we're a little bit low on time but there's a lot of ways that um you know our access product is able to do this kind of this authentication and asserting different policies on who's able to access what is it the access application contains kind of a client id and a client secret and so what that allows us to do is scope it down as narrowly as possible right a lot of times we talk to customers they may have these like wide kind of authentication policies or authorization policies on who's able

13:01to access what but given that we have an access or access product you're actually write pretty narrow policy groups that can go in and say, for this user, these set of users, I want to be able to have these set of actions or be able to authorize them to view certain things, right? And really that lets you have more control over who's accessing what data and being able to see in the agent realm, right? How the agent that that person is using is also what they're able to authenticate into and authorize their application or authorize their access to different applications. And in this case, we'd like to make sure that the access has both credentials, right? We want to make sure tokens, I think, when a lot of these things kind of initially spun up, you know, people were using service tokens, and we found that can work, right?

13:51But it does kind of open you up to quite a bit of risk, especially with people not scoping their tokens properly. They have really long, you know, times that the token... can continue to exist and then now you have a lot of different problems that can come up especially if there's any sort of token theft right that can happen suddenly you're on the hook uh it just happened with open ai you know or sorry open code uh where you know people were able to steal tokens and charge other people's accounts right so that doesn't happen if you're using uh access and authentication that that allows you to to do these tunneling and to access these different applications and apis so uh the the main way we kind of expose things to the internet is using Cloudflare tunnel. It's Cloudflare D if you ever run it in any of your boxes.

14:37And what that allows you to do is kind of have a tunnel out to these different private APIs, and then you can put access on it to be able to have those different authentication policies. And again, your access kind of authorizes. And then once it's authorized, the tunnel will then act as that transport layer between the data between the private API and whatever agent that you're running itself. This is kind of the way that we set up that lab, but also how we recommend most customers kind of set this up as well with their agent. And then to go to the second half of that question, which is, hey, is this data safe, right? Is the body inspection, right? And so when you're having data grow across our gateway, obviously we're decrypting that with TLS. And then we're actually running that against different checkers, right? That may be your DLP policies if you want to check for sensitive information.

15:25You can also... make sure uh if there's any prompt injection attacks that are there we do have that capability as well so us being able to inspect that traffic really gives us that insight but you as kind of the end user to be able to write different policies and control what is happening with that agent and how they're interacting with excuse me the control endpoint or application or wherever that agent is going so that is one of the ways again that was the last uh example that we had showed uh a large kind of group there so again i'm gonna summarize this in a little bow before i hand it back to todd on kind of more concrete ways on we've seen customers kind of deploy this is what we verified here is we can look at different policies around blocking you know sensitive information we can make sure that the data that the agent is is going through it's going through our gateway it's going through those checkers and we can decrypt it and also we can check for authentication and make sure that using access that that agent is authorized to access those different applications without needing to use static service tokens.

16:31And again, as you evaluate within your pilots or if you're running your own agents, really look at different LLM and prompt behavior attacks. We see that a lot. Todd and I come from a background in email. We're seeing it now even in email or phishing attacks. know, are whiting out parts of the message, and then they're hoping that it'll be a prompt injection attack, right? Because agents read emails all the time now, right? So really something that we've seen a lot of customers either get hit with that we can kind of protect against, and then just really focusing on credential revocation and failover as well. So a lot of things here, but I'm going to hand it over to Todd to talk through some of the ways that he has seen how to actually do this in production as he's seen.

17:18different customers yeah so there's a lot of new tech obviously some things still work great right pki is always going to be a pain in the butt however it's functional and it does a great job for what it does as long as you manage their certificates um the big thing is rotation is important i think that's a step everybody forgets they forget it with the access tokens they forget it with pk asserts right we all kind of edge to the lazy side say yeah five years is fine right it probably isn't but the benefit here is you don't have to spend five years on this stuff you can you can build it with terraform you can automate this thing this you can use ansible or any other technologies get yourself to shortly short -lived certificates right maybe you use our pki or somebody else's pki and you do daily certificates you just simply rotate them off but that's the first part right so we have to secure where the device is coming from next week i'll look rotation the expiration everybody see the news about the intermediate cert problem today that's not nearly enough hands so there's a big intermediate cert i think there was five or six of them um the the intermediate ca is where were uh yeah the key was four uh forged right it was yeah needless to say you've got five or six intermediate trusted certificate authorities that are now issuing bad certs right so this stuff has to be managed

18:49fortunately we can do a lot with AI really really fast right so what I did at least in this case is I picked the workload right my my goal here was a private internal resource that could be something that's connected to a SaaS service it could be you know in the healthcare world maybe it's epic maybe it's some kind of certificate database maybe it's your applications maybe it's your rams right whatever the case pick something that you actually need to secure but try and start with some place where you're actually trying to enable people to work right the goal here is make sure that the communication is secure it's coming from the right devices and the responses coming back that might be going to ai are actually appropriate right from there you can define the actual destinations what data should be in there this is very typical old style dlp functionality right if you don't understand the flows and the data classification how do you build policies from there you create your tests we make it really really simple access is very very good at this stuff you put tunnel in front of the private resource you create the access token you designate the jwt for the certificate that's about all there is to it right it's a handful of policies and probably 20 minutes worth of effort From there, you can actually test it.

20:09You can send some test cases through the APIs, you can send some bad responses back, see what happens, tune your product, and then you can start to look for your production side. So again, automate your credentials, automate the PKI, go fix those problems first. Make sure you have resilient connectivity. If you're gonna use Cloudflare D, you want multiple replicas, right? Everybody familiar with that? Anybody deployed it before? Okay, so I suggest at least two, right? As a pair of replicas, you can go further if you want, but for the most part, it's gonna handle a ton of traffic and a lot of requests, but get the connectivity side. From there, you can go onto the data policies, make sure that you know what data is appropriate to come out and the build your policies to interact with that.

20:57Consider accounts, consider things like exact data matching, It's a wonderful technology to be able to say, here's the stuff that I want to block for sure and if it's not in that data set then maybe I I use regex or something less deterministic and then lastly take that gather the data look at the logs make sure it's appropriate from end to end and then keep going all right and here's a quick look at kind of what the demo lab looks like from an architectural standpoint there's also a link to cloudfire one page if you're already using say one of our internal bundles or access or gateway you already own this stuff right most of this are tools that are accessible to you today right if you're a small shop as much as I dislike to say it but our free plan gives you a ton of tech you would not necessarily need to pay for it in some cases right but I work in sales I didn't say that cool uh thank you all for coming we're a bit over time but we'll be here outside, feel free to reach out to Todd and I.

21:59We're here for any questions, any way we can help you set up whatever problems you're running into. We're always here as a resource. So thank you all again for taking the time to come out to our talk. Any questions before we go?

22:15No? No. Okay. We'll be outside if you have anything later, but thank you all again. And have a great rest of your Connect conference. Thank you all. Thank you.

Browse all
Connect 2027

Connect returns in 2027

Be the first to know when registration opens and the agenda goes live.

By submitting this form, you agree to our Privacy Policy.

cloudflare-connect-2026

---
title: "Protecting AI agents with Cloudflare One"
url: "https://www.cloudflare.com/connect/2026/session/protecting-ai-agents-with-cloudflare-one"
year: 2026
---

# Protecting AI agents with Cloudflare One

[Watch session](https://www.cloudflare.com/connect/2026/session/protecting-ai-agents-with-cloudflare-one)

## Session details

- **Track:** Secure AI
- **Topics:** Agentic identity, Data security & privacy, Agentic AI & MCP
- **Solution areas:** SASE & Workspace Security (Cloudflare One)

## About this session

AI agents are proliferating inside enterprise networks with static API keys and no identity checks. This session deconstructs how Cloudflare One treats agents as first-class identities at the edge.  Explore machine-to-machine authentication via Access service tokens, short-lived mTLS, and ZTNA segmentation enforcing least privilege without network complexity. Also, learn how you can use Cloudflare Gateway to inspect agent-to-SaaS traffic in real time and DLP policies to prevent exfiltration. The outcome: a concrete architecture that replaces secret sprawl with policy-driven enforcement.

## Speakers

- [Ayush Kumar](https://www.cloudflare.com/connect/speakers/2421441) — Principal Product Manager, Cloudflare
- [Todd Murray](https://www.cloudflare.com/connect/speakers/2453475) — Solutions Engineering Manager, Cloudflare

## Transcript

[0:02] All righty. Hi, y 'all. My name is Ayush. I'm joined here with Todd. This is the Protecting AI Agents with Cloudflare One. It's going to be about 20 minutes. We do have a live demo and a little bit more hands -on stuff that we'll show you on how to do this. So strap in. If you have any questions towards the end, we'll be here afterwards as well. So please, you know, save them to the end because we have a little bit of material to get through. Before we I'd like to introduce ourselves. I'm a Principal Product Manager at Cloudflare. I deal with a variety of things, but specifically thinking a lot about AI agents and how we do behavioral analysis and things along that side for them. Yep, Todd Murray. I manage a group of solutions engineers who are out of the Americas. We're responsible for AI, Zero Trust, SASE, email security, things along those lines.

[0:51] So, perfect. What you're here for is we wanted to kind of start off with an example of what you may be running into within your organizations is you set up an agent that's supposed to read an internal record it's supposed to prepare a summary and send it to an external endpoint right a very common agent probably one that you may be running already internally or looking to actually deploy but what happens if the agent accidentally sends the hold record right it's supposed to only send a small part or summary what is going to happen if it accidentally sends the whole thing. We talk a lot with customers that are really worried about the fact that these agents are non-deterministic. It's really hard to know what their expected behavior is going to happen, right? And so how do you prevent this agent from going off the rails by providing it some of the guardrails?

[1:38] And this is kind of what we're going to be talking about in kind of the demo.

[1:43] So we see a lot of this when we talk to customers and how they're kind of deploying it themselves. For example, oftentimes they'll use a secret, they'll use a token, right? These tokens can... sometimes kind of outlive the tasks that they're used for, right? So you have an agent that has a token, it was only supposed to do it for a short amount of time, but now this token has continued to exist and maybe misused or the agent can continue to use it over and over again in different ways. You know, you have the network access becomes application trust. You have the full record leaves unchecked and then decision -like usable evidence. So kind of going through those things, what those actually mean is It's really hard to know, for example, when you're decrypting maybe like a website, right? You're trying to look at kind of traffic going through your network.

[2:30] You have to do, you know, all these body inspections or requires TLS decryption. But it can still, if you have an authorized read, it can still lead to kind of an unsafe kind of upload, right? So how do you prevent your agent from doing things that it's not supposed to be doing and to make sure that you've really scoped it very cleanly? to the task that it's supposed to be doing so we kind of think of it in two questions and kind of two controls here so what way what may my agent reach and what may my agent release right and so the way we split this up and we'll see we'll see more and actually the demo itself is on one thing on one side the first one what may my agent reach is really the authentication and authorization problem right what is it authorized to read what is it authorized to actually view what is the data access to, right? And the second one is what made the agent release, right?

[3:16] Do you want your agent to be able to send out PII information, right? Do you want them to be able to send out sensitive social security numbers that it may be reading from somewhere or PII information like health records, right? We see this often, right? And the way that we conceptually like to talk about with our customers is these are two different kind of problems, right, that you're trying to deal with when you're building your agents. And we'll kind of get into the products So we have our access and kind of gateway DLP product that helps those two different things. So we're actually going to do this demo. We're going to show you a live demo of this. The first one, excuse me, is a headless agent. And it's going to go through our Cloudflare Edge using our Cloudflare tunnels or Cloudflare D. And then we're actually going to hit a private application like a private API.

[4:05] Excuse me. And the second one is we're going to use our Cloudflare 1 client, which is our warp client. And that's actually going to proxy the traffic. through the Cloudflare Edge, using Gateway and our DLP. And this is the second question, right? How do you control what's going out there? What can we kind of analyze, whether that's the body, what the contents of the request is? And then we're going to go to a controlled external endpoint. So these are the two things we're going to do a quick demo on. I'm going to hand it over to Todd. We're going to walk through his setup and how these things actually work like in practice and how you can set it up yourself at home. So I'm going to show you the... demo, but first let me kind of describe the lab. I built a lot of these. This one in particular is Proxmox hosted inside of GCP so they can do layered virtualization, right?

[4:53] From there, effectively, it's designed to look like five or six different data centers. It's all interconnected with SASE. You can see a whole host of routers down at the bottom. All those are running our own Cloudflare WAN technologies for connectivity. Then on top of that, I've got some mcp services to manage the proxmox environment do a lot of the builds i've got some clients to actually reach out and and send data back and forth and then finally you've got the demo client there and then the one down here and those are the two that are kind of a deterministic state that i put in there in order to send data back and forth right so that's what we're going to focus in on um for fun and i'll show this at the end everything i built i built the AI. I've also put it in our internal GitLab, but I took it and I also published it out to GitHub if you wanted to basically duplicate this as a test bed.

[5:41] So you can set it up, it'll walk you through all the steps to setting up access and MTLS, perform all the security elements of it, put it in the live lab, you can test it, see the data, the requests, and then adopt that to production use, right? And we'll go through that. But here's kind of the page that we put together. I'll show you the actual logs, but what it's going to do is it's going to intercept those things. We're just going to look real quick at the different use cases, right? So the first is how do we establish identity for the client? In this case, we're mostly doing it around MTLS, right? So we're going to use certificates to validate that the machine should be able to have communication. We're using private infrastructure, right? So PKI is built into the demo.

[6:27] And there you can just kind of see there's the request. If we're not sending anything, we're just verifying the actual... issuance. All right. On the next, we're going to send a request and we're going to do this without authentication. Right. As you would expect, it's not going to get where it's supposed to be going. Right. We're going to block it. There's going to be no response whatsoever. All right. Now, tokens are great. They're a great approach. What they don't do is necessarily validate the origin server. Right. They validate that the request came from somebody that it probably should, but you still don't know whether or not it came from the right infrastructure. So if I separated my machine from a certain agent, for example, right? You've heard about Cloudflare OS. If Cloudflare OS was talking to an internal service, how do I separate it from me?

[7:12] This is how we would do that is on the TLS side, right? So token only, same thing, we send the request, we send the token, we know it's valid. We're still not gonna get a good response, right? Take the next step. Here we're gonna send both the token we're going to validate the certificate as part of the actual request all right so there's your get and now we're actually going to get back to 200 which means those two services are going to be able to talk correctly so we've verified the inbound direction we now know it's coming from the right service we have token for authorization but we have tls to verify the actual origin all right so next we're going to send data make sure that goes through great all right we can post the data out so we push stuff in right so this is our response going out to a service this could be an mcp it could be a sas service whatever whatever that tooling is but this is your outbound right and then the last step is let's look at the response side the problem here and this is why it's kind of important is because when in the lm world you don't necessarily know what that AI agent is doing. It's doing all its work. It's going to skim past.

[8:25] You're going to see some logs. You don't have any time to interrogate that. You certainly don't have time to look at the response and interrogate that to see if what is coming back to the AI from those tools is appropriate. Is it the right data? Is it too much data? Did I just cause an incident or a breach because I pulled, you know, instead of a first name, I just pulled somebody's entire health record. right it's really easy to make that mistake you might not find out until you're in the news because suddenly your data is showing up in a generative AI right that's what we're trying to avoid here so in this case what we're doing is we're taking the authorization now we're on the response coming back from the tool before it goes back out to AI and we're able to actually look what that data is and block that from a DLP perspective right no human interaction is involved in any of this

[9:23] right now on this if you do want to take it and use it feel free I just pushed this over today it should work I'm gonna go with should because I didn't test it but effectively what I did is I took this demo lab I put it up at github so that you could try it and use it up right

[9:49] On the log side, mine is light in there, but this is just to kind of show you're going to see the same stuff on the log side, right? It's going to come in the clefler axis. You can see the actual requests, whether or not they're forbidden or whether or not they're allowed. Here you've got my OKs, right? The actual JSON requests behind those payloads. So the proof is pretty easy to get to in that case. All right, like I said, I built this all. off of ai so it's easy to rebuild and then it is published out so if you didn't get the qr code it's just mellow techie it's kind of my handle is where all my sites and my demo domains look but you'll just find it right there okay back to the deck please yes black back to the deck

[10:39] cool so you know we had 20 minutes so we kind of blazed through that a little bit but really to go back to the core idea of this is that we just showed you five different outcomes, something that, you know, your agents will be doing maybe in production, maybe just an internal capacity. But the first one we showed was with no credential, obviously very simple to do, right? Access denied. But the next one, especially as you're hitting these private API endpoints, whether that's an internal server, or you're running that you want your AI agent to maybe get health stats for, it could be a plethora of things we've heard our customers talk about. You can see that just having a token, you can an access deny, the API denied it. But if you want to make sure with MTLS that you have both the token and the valid certificates, that they can have that conversation, you can do so there. And you saw that in kind of the third example. And then finally, a safe post, right, which is, hey, I'm querying my Salesforce, right, and I want to pull some information from it.

[11:29] But I want to make sure that, you know, my agent is not pulling sensitive information about that Salesforce record and then feeding it. to a generative AI that, or a generative LLM that now is being trained on that data, we showed in one way, you know, without any sort of sensitive information, it went through, but then in our ability to block that by inspecting the traffic, we were able to show that in the end that we were able to see, hey, this had some PII, some, you know, information that should not be leaving based on a policy and we were able to block it, right? So that was kind of the big TLDR is like, look, we gave you some examples, obviously you have Todd's. of uh test test uh github that you can do and spin it up and play around with yourself when using tunnels using access using gateway and using dlp and all these different building blocks to kind of play around with it yourself on on different uh use cases that you have um and so kind of how a little bit deeper uh and we're going to kind of go through a little bit fast as well because we're a little bit low on time but there's a lot of ways that um you know our access product is able to do this kind of this authentication and asserting different policies on who's able to access what is it the access application contains kind of a client id and a client secret and so what that allows us to do is scope it down as narrowly as possible right a lot of times we talk to customers they may have these like wide kind of authentication policies or authorization policies on who's able

[13:01] to access what but given that we have an access or access product you're actually write pretty narrow policy groups that can go in and say, for this user, these set of users, I want to be able to have these set of actions or be able to authorize them to view certain things, right? And really that lets you have more control over who's accessing what data and being able to see in the agent realm, right? How the agent that that person is using is also what they're able to authenticate into and authorize their application or authorize their access to different applications. And in this case, we'd like to make sure that the access has both credentials, right? We want to make sure tokens, I think, when a lot of these things kind of initially spun up, you know, people were using service tokens, and we found that can work, right?

[13:51] But it does kind of open you up to quite a bit of risk, especially with people not scoping their tokens properly. They have really long, you know, times that the token... can continue to exist and then now you have a lot of different problems that can come up especially if there's any sort of token theft right that can happen suddenly you're on the hook uh it just happened with open ai you know or sorry open code uh where you know people were able to steal tokens and charge other people's accounts right so that doesn't happen if you're using uh access and authentication that that allows you to to do these tunneling and to access these different applications and apis so uh the the main way we kind of expose things to the internet is using Cloudflare tunnel. It's Cloudflare D if you ever run it in any of your boxes.

[14:37] And what that allows you to do is kind of have a tunnel out to these different private APIs, and then you can put access on it to be able to have those different authentication policies. And again, your access kind of authorizes. And then once it's authorized, the tunnel will then act as that transport layer between the data between the private API and whatever agent that you're running itself. This is kind of the way that we set up that lab, but also how we recommend most customers kind of set this up as well with their agent. And then to go to the second half of that question, which is, hey, is this data safe, right? Is the body inspection, right? And so when you're having data grow across our gateway, obviously we're decrypting that with TLS. And then we're actually running that against different checkers, right? That may be your DLP policies if you want to check for sensitive information.

[15:25] You can also... make sure uh if there's any prompt injection attacks that are there we do have that capability as well so us being able to inspect that traffic really gives us that insight but you as kind of the end user to be able to write different policies and control what is happening with that agent and how they're interacting with excuse me the control endpoint or application or wherever that agent is going so that is one of the ways again that was the last uh example that we had showed uh a large kind of group there so again i'm gonna summarize this in a little bow before i hand it back to todd on kind of more concrete ways on we've seen customers kind of deploy this is what we verified here is we can look at different policies around blocking you know sensitive information we can make sure that the data that the agent is is going through it's going through our gateway it's going through those checkers and we can decrypt it and also we can check for authentication and make sure that using access that that agent is authorized to access those different applications without needing to use static service tokens.

[16:31] And again, as you evaluate within your pilots or if you're running your own agents, really look at different LLM and prompt behavior attacks. We see that a lot. Todd and I come from a background in email. We're seeing it now even in email or phishing attacks. know, are whiting out parts of the message, and then they're hoping that it'll be a prompt injection attack, right? Because agents read emails all the time now, right? So really something that we've seen a lot of customers either get hit with that we can kind of protect against, and then just really focusing on credential revocation and failover as well. So a lot of things here, but I'm going to hand it over to Todd to talk through some of the ways that he has seen how to actually do this in production as he's seen.

[17:18] different customers yeah so there's a lot of new tech obviously some things still work great right pki is always going to be a pain in the butt however it's functional and it does a great job for what it does as long as you manage their certificates um the big thing is rotation is important i think that's a step everybody forgets they forget it with the access tokens they forget it with pk asserts right we all kind of edge to the lazy side say yeah five years is fine right it probably isn't but the benefit here is you don't have to spend five years on this stuff you can you can build it with terraform you can automate this thing this you can use ansible or any other technologies get yourself to shortly short -lived certificates right maybe you use our pki or somebody else's pki and you do daily certificates you just simply rotate them off but that's the first part right so we have to secure where the device is coming from next week i'll look rotation the expiration everybody see the news about the intermediate cert problem today that's not nearly enough hands so there's a big intermediate cert i think there was five or six of them um the the intermediate ca is where were uh yeah the key was four uh forged right it was yeah needless to say you've got five or six intermediate trusted certificate authorities that are now issuing bad certs right so this stuff has to be managed

[18:49] fortunately we can do a lot with AI really really fast right so what I did at least in this case is I picked the workload right my my goal here was a private internal resource that could be something that's connected to a SaaS service it could be you know in the healthcare world maybe it's epic maybe it's some kind of certificate database maybe it's your applications maybe it's your rams right whatever the case pick something that you actually need to secure but try and start with some place where you're actually trying to enable people to work right the goal here is make sure that the communication is secure it's coming from the right devices and the responses coming back that might be going to ai are actually appropriate right from there you can define the actual destinations what data should be in there this is very typical old style dlp functionality right if you don't understand the flows and the data classification how do you build policies from there you create your tests we make it really really simple access is very very good at this stuff you put tunnel in front of the private resource you create the access token you designate the jwt for the certificate that's about all there is to it right it's a handful of policies and probably 20 minutes worth of effort From there, you can actually test it.

[20:09] You can send some test cases through the APIs, you can send some bad responses back, see what happens, tune your product, and then you can start to look for your production side. So again, automate your credentials, automate the PKI, go fix those problems first. Make sure you have resilient connectivity. If you're gonna use Cloudflare D, you want multiple replicas, right? Everybody familiar with that? Anybody deployed it before? Okay, so I suggest at least two, right? As a pair of replicas, you can go further if you want, but for the most part, it's gonna handle a ton of traffic and a lot of requests, but get the connectivity side. From there, you can go onto the data policies, make sure that you know what data is appropriate to come out and the build your policies to interact with that.

[20:57] Consider accounts, consider things like exact data matching, It's a wonderful technology to be able to say, here's the stuff that I want to block for sure and if it's not in that data set then maybe I I use regex or something less deterministic and then lastly take that gather the data look at the logs make sure it's appropriate from end to end and then keep going all right and here's a quick look at kind of what the demo lab looks like from an architectural standpoint there's also a link to cloudfire one page if you're already using say one of our internal bundles or access or gateway you already own this stuff right most of this are tools that are accessible to you today right if you're a small shop as much as I dislike to say it but our free plan gives you a ton of tech you would not necessarily need to pay for it in some cases right but I work in sales I didn't say that cool uh thank you all for coming we're a bit over time but we'll be here outside, feel free to reach out to Todd and I.

[21:59] We're here for any questions, any way we can help you set up whatever problems you're running into. We're always here as a resource. So thank you all again for taking the time to come out to our talk. Any questions before we go?

[22:15] No? No. Okay. We'll be outside if you have anything later, but thank you all again. And have a great rest of your Connect conference. Thank you all. Thank you.

## Related sessions

- [Securing agentic harnesses: Control, visibility, and safe tool use](https://www.cloudflare.com/connect/2026/session/securing-agentic-harnesses-control-visibility-and-safe-tool-use)
- [Secure AI at scale: Cloudflare’s vision for trusted AI adoption](https://www.cloudflare.com/connect/2026/session/secure-ai-at-scale-cloudflares-vision-for-trusted-ai-adoption)
- [After human: How to think about fraud & abuse when anyone could be anything](https://www.cloudflare.com/connect/2026/session/after-human-how-to-think-about-fraud-and-abuse-when-anyone-could-be-anything)

## More from Connect

- [All sessions](https://www.cloudflare.com/connect/)
- [Get notified for Connect 2027](https://www.cloudflare.com/connect#2027)