Governing Enterprise AI at Scale: Zscaler + Anthropic

  • Zscaler
Governing Enterprise AI at Scale: Zscaler + Anthropic

About This Webinar

Enterprise AI is moving into everyday work. The challenge is no longer access - it's control.

Join Zscaler and Anthropic to explore how organizations can bring security and governance to AI in production. Hear from Dor Fledel, Anthropic, on how Inference Hooks enable inspection of prompts, uploaded files, and tool-call responses before they reach the model. Then see how Ashwin Kesireddy, Zscaler, explains how Zscaler Security for AI applies inline DLP, intent detection, and content policy enforcement to help teams reduce risk and move faster with confidence. You'll also hear Patrick Abreo, PayPal, on what governed AI deployment looks like in practice.

Hear how leaders from Anthropic, Zscaler, and PayPal are delivering AI governance in the real world - and what your team can do next. If governing AI at scale is on your roadmap, this is a session you won't want to miss.

What You'll Learn

  1. How Anthropic Inference Hooks expand visibility into AI interactions
  2. How Zscaler Security for AI helps enforce inline controls
  3. How PayPal is approaching governed AI deployment

Speakers

  1. Ashwin Kesireddy VP, Product Management — Security for AI Zscaler
  2. Dor Fledel Inference Hooks & AI Governance Anthropic
  3. Patrick Abreo Director, Cybersecurity Engineering PayPal
Full Transcript Scott Bekker Hello, and welcome to Governing Enterprise AI at Scale, Zscaler, and Anthropic. Today's webinar is sponsored by Zscaler and, produced by Actual Tech Media. My name is Scott Becker with Actual Tech, and I'm excited to be your moderator for this webinar. So we've got, great stuff to cover today. Can't wait to, get into it. But first, do wanna go over a few housekeeping details that are gonna help you get the most out of this session. So, three tabs over on the right I wanna call your attention to. There's a chat tab, a docs tab, and a q and a tab. The chat tab is, with the one a lot of you have already found. You're already, saying hello in there. That's great to see you. We got somebody from Idaho, Wisconsin, Arizona, California, Bethlehem, Atlanta, California. Adam is back. Welcome back, Adam. Colorado and, Montana, great, great spread of states there. Thanks everybody for joining, and, please keep those, keep those hellos coming. We'd love to see where you're logging in from, and, of course, we'd like to see, you know, more than just your hellos, you know, just your your comments on on the, you know, what you're hearing during the presentation and, and, you know, any reactions that you have and and, thoughts you wanna share with each other. It's, that's a perfect place for it. We got some people clocking in from, Michigan and Maryland as well. Awesome. So, also, we've got a, the middle tab. Next to the chat tab is the docs tab, and, we've got some good stuff in there too. There… There's only one link, to a a res… A Zscaler resource page, But that includes things like, Zscaler's Threat Labs 2026 ransomware report, really cool resource. Some resources about how the the Zscaler platform beats AI threats. There's case studies on there. There's a place to request a demo. So all kinds of, different things on that on that page. It's a great page to go to. And then the third tab is the q and a tab. And, I don't think we're gonna have a a live q and a session on this one, but any questions that you have for Zscaler, or just for an… Anthropic, you can you can drop there in that q and a, and we'll pass that along, to the team there for, for follow-up. Now also be aware that we have clap… Closed captioning available in 18 languages. If you just hover over the lower part of your screen, you'll see that CC option as well as a drop down menu for languages. Finally, you may have noticed the prize icon. We are giving away an Amazon gift card worth $250 as a prize drawing at the end of our webinar. Now, you gotta be in attendance for the duration of the live event to qualify for the prize. You gotta meet the actual tech media prize terms and conditions, which you can find in the docs tab, but that's a that's a great way to have, end of Friday is, get that, that gift card, and then you've got Black Friday coming up in just a week. So it'd be very timely. Wish I was in the running for that that gift card. Alright. With that housekeeping out of the way, let's get into our session and, really excited to introduce you all to this… The speakers that we've got, lined up for you today. So we've got Ashwin Casaretti, who's a vice president for product management of security for AI at Zscaler. And he's gonna be kind of, you know, presenting, but also emceeing, the event. Then we've got Dorf Fladell, who's Inference Hooks and AI Governance at Anthropic. And we have, Patrick Abreu, who's director of cybersecurity engineering at PayPal. So great lineup. Let's, get started without further ado. Video Hello, everyone. Welcome to this event talking about AI security solutions. At Zscaler, we are helping our customers adopt AI at scale, and we are working together with ecosystem partners to achieve business outcomes for our customers. My name is Ashwin Kesriedi. I'm the VP of Product Management, AI Security at Zscaler. Joining me is a special guest from Anthropic, Dor Fleddle. Dor? Hey, everyone. I'm Dor from Anthropic, leading the product on security and enterprise readiness. Really excited to be here. For the audience joining this event, it should not be a surprise that over the last year, AI has seen dramatic growth across all facets of our professional lives. We have seen AI getting embedded in various SaaS applications, workflows, tasks, processes, you name it. Now this dramatic growth of AI has led to a significant growth in agents as well. We have seen agents sprawl happening everywhere, both on the endpoint as well as in the cloud. This naturally raised a lot of questions about things like agentic identity, nonhuman identity that you assign to agents, about agent authorization, as well as about the intent of these agents as these agents are taking various actions, both on the laptop and in the cloud. This is naturally creating a lot of governance issues for our customers. And as part of this governance, customers are looking for visibility into different agentic actions. They're looking for granular controls as well as operationalization of these controls. These governance requirements are not specific to any channel. There's spread across web, email, cloud, you name it. And customers also have to worry about managed and unmanaged devices as well as authenticated and unauthenticated traffic as well. Dor, would you like to share some insights here? Yeah, Ashwin. I think that really aligns with what what we're seeing. I think as AI evolution and maturity grows, as customers are moving to longer horizon tasks and more agentic flows, it really aligns with full coverage than they want, across their environment. And I would add to that that we're seeing slightly different scrutiny around developers and coders use cases versus what we call knowledge workers and, and regular employees, flows where security teams want to take slightly more specific controls with each ones according to their threat moves. Addressing these customer challenges, Zscaler has been delivering several innovations focused on AI security. Few years ago, we started off with the hardest problem in AI security. That is the AI guardrails product. We started discovering shadow AI usage in our customers' environment. Last year, we expanded that to security posture management, AI asset management, AI red teaming. And this year, we have added AI access graph, endpoint AI security, AI gateway. So the pace of innovation just continues to accelerate as we are solving many use cases and problems for our customers. In this event, we are particularly focused on anthropic inference hooks, and the product that is very close to anthropic inference hooks is Zscaler AI Guardrails. As part of Zscaler AI Guard, we have several detectors that act on the prompt and the responses. These detectors act as content filters. They act as your security gateway keepers to protect from prompt injection attacks. They protect leakage of sensitive information or any custom topic you define. Detectors act on the prompts and make sure that your information is secure and your agents are secure. Now Zscaler has been a man in the middle proxy. We can look at both prompt and the response. So these 20 plus guardrails apply not only on the prompt but also on the response side. This gives customers most comprehensive prediction from a prompt response guardrails perspective. We delivered guardrails primarily for our Zscaler Internet access or ZIA customers in… At the beginning. Recently, we expanded that to ZPA or Zscaler private access customers as well. And very recently, we also introduced the concept of bring your own detectors that let customers flexibly consume various guardrails from the Zscaler platform itself. Now Zscaler's innovations are not just limited to our core products. We are partnering with ecosystem vendors as well as frontier models like Anthropic. And particularly with Anthropic, we are work together with them to deliver anthropic inference hooks. Anthropic inference hooks captures prompts from the users. And no matter where the prompt is entered, whether it's in Cloud Code, Cowork, or Cloud dot AI, prompt gets captured, and Anthropic inference hooks makes an API call to… Authenticated API call to Zscaler AI Guard. Zscaler AI Guard sends a response back to Inference Hook whether the prompt is benign or not. If it's benign, the prompt is allowed, and it gets processed by cloud. If it's malicious or if it has some sensitive information, Zscaler AI Guard sends a block response back to anthropic, and anthropic inference hooks blocks the prompt from being processed by the cloud models. Now anthropic inference hooks, in addition to working only as allow and block mode, they also have a shadow mode that lets organizations, detect events rather than acting on the events. This gives lot of flexibility to customers to consume AI Guard. So you can consume the AI Guard from ZIA or ZPA or from anthropic inferences itself, giving customers more options on how they, deploy Zscaler AI Guard. I would like to bring Dor back into the conversation to see if it has any more insights here. Hey, Ashwin. Yeah. Everything that you said makes a lot of sense, and I'm really excited about how we developed, this product and really about the the value to shared customers. I think that in general, I would love to highlight two main things. One is completeness. So one of the things that we're really getting from the cloud hosted hook that we've created, and the customers are… Some of them are already familiar with the the cloud code hooks, which we kind of took as inspiration here. So the cloud hosted solution provides completeness. Every prompt that's sent by any one of your users will get hooked, with the only single deployment across your your infrastructure. And that's very powerful for security teams, especially the ones that want to get from, further in their AI security maturity. The second thing that customers really, really told us is about the power of InAlign. I think now as AI adoption grows, as you… There is need to get more educated about AI, how it works. It's more powerful by its nature. The inline validation, which is some areas of security didn't exist in the past, is very key to our customers as they get more certainty and more controls about the data inflow and outflow from their organization and get a very positive feedback. Now I would like to show you how Anthropic Inference Hooks and Zscaler AI Guard work in action. I will start this demo by entering a benign prompt into my Cloud chat. This prompt does not have any sensitive information in it, and it's a pretty simple task of composing an email. The prompt goes through because it's very benign, and a response comes back. Now I modify the prompt by including sensitive information, sensitive PII information. In this case, I have included some dummy credit card information. You'll see that Anthropic Inference checks kicks in and blocks the prompt. What you are seeing on the screen is also a custom block message that you can set up on the Inference page. As a savvy insider, I could take that prompt with sensitive information and put that in a document And ask Claude to follow my instructions. You will see that anthropic inference against kicks in and blocks the prompt. What is happening in the background is Anthropic Inference Hooks is making an API call to Zscaler AI Guard, which is sending a block signal to the Inference Hooks, which enforces the block. And on Zscaler AI Guard dashboard, we capture all the allowed and blocked transactions. So far, I have demonstrated blocking of PII information. Now, I am demonstrating blocking of custom topics using Anthropic Inference Hooks. For the sake of this demo, I set up a custom topic on Zscaler AI Guard to block discussions about compensations of colleagues in a company. You'll see that Anthropic Inference works again kicks in and blocks the prompt. This transaction is again captured on the Zscaler AI guardrail dashboard, and you will see that the topic detector got triggered, and you can also view details of the transaction. Working together with Anthropic, we have started off on the prompt side, but there are more new exciting announcements coming out as part of this Inference Hooks work that we are doing with Anthropic. To talk more about this, let me bring Dor back into the conversation. Dor, take it away. Yeah. For sure. So we released this innovative capability around two months ago. And now as customers… More and more customers are are using it, we're getting amazing feedback. They're really excited, and there is more to come. So we started with a very immediate prompts and inference. Whatever gets to the inference gets hooked. That's Ashwin's prompt that's, additional things as they go to the inference. We're adding more and more coverage as we speak. The next thing which we added, early October just last week was tool call. So examples for tool calls are one MCP tool call. So before your calendar query goes off or similar, that's being looked. Another one which is very relevant from other customer conversations are web searches. A lot of customers are worried that their customers or clients or users information would be released in a web search, and they want us to have them hook it. And that all was delivered already. Same configuration, just one more toggle, for for admins. So very easy deployment, and customers already gave us very positive feedback. And, what's next? We're constantly expanding the coverage and and listening to customers who are getting feedback about content redaction. So not blocking the prompt entirely rather than just removing certain certain parts of it. We're adding more, information about files and uploads and additional things. And, we have other more ideas, but in general, we want to make deployment as simple as possible. We already have the shadow mode and percentage rollout in as well as exclusion groups, and we want to make it as easy as possible. So please, tunnel your feedback, into the zscaler or or our team because that feedback for us is gold at this point, and we really are bullish on on the approach that we have here together. Wow. That that was amazing. I know we are working on more innovations together. I appreciate you participating in this event with us. Now to continue this conversation and get more insights, we have a special guest joining us. We have invited one of Zscaler's customers who has been on this journey with us since the beginning. Please welcome Patrick Abreu from PayPal. Patrick, thank you so much for taking the time. Now everybody knows PayPal, global payments company, global organization. But just to get started, can you share a quick high level overview of your responsibilities at PayPal? Absolutely. And thank you for having me here. It's it's a it's a pleasure to be here on this particular call. So I lead workforce identity and access management at PayPal, including network access. And my team's role is to make sure that the right people, including agents now, have access to the right resources at the right time. And this makes me and my team at the at the focus of securing access to agents and LLMs. That's become the core business point of my lead and my job at PayPal. Patrick, I think it would be really helpful if you could touch on the use cases and business outcomes you are driving with AI. Can you share how Zscaler AI guardrails, helped you in the process of driving business outcomes? Absolutely. The… When AI was first introduced here at here at PayPal, we had access to a broad set of LLMs and and AI agents. Now this was rather rather quick. We went from zero to 100 as isn't the traditional PayPal case within a very short period of time. That was great, provided great productivity, improvements to the business, speed of execution improved, the quality of work improved, and everyone was happy. However, that came with a challenge. The challenge was that we need… We lack visibility. What were users entering into the LLMs, and what was the output that was being generated? We needed to make sure that there is access to LLMs was securely enabled and that we had visibility into the prompts and outcomes so that our folks were guarded, our business was guarded, and the route outcomes were delivered to the business. This is where Zscaler, along with your… The AI guardrails, helped us to get that visibility into what was required and securely enable the business. Love to hear that. Now your team at PayPal has been very creative in leveraging ZIA and ZPA for AI security. Can you touch on, the different ways you're using these services to enable AI security in our organization? Absolutely. So just taking a step back and and talking about our journey with Zscaler. We always had ZIA in our organization for a while in order to provide us that as secure web gateway slash SASE interface for access to external websites, tools, providing a data less prevention, and also in making sure that we were guarded from all cyberattacks. Then we went ahead and deployed ZPA on top of it so that our access to our internal applications was was following a zero trust network access path. So we had both of these p… Pieces already in place, and where an AI gate card rings came in, it was a flip of a switch. It was completely transparent across to our users, and we could deploy it in a very short period of time. And it provided us the ability to tune it over a period of time, which was also great. Where we are gonna take this from here is now this is user access to relevance. You also want to act, secure the access from our MCPs and our internal LMS to external LMS. That's where we're gonna take it forward with Zscaler's detector as a service architecture. Thank you. Excellent. Now this webinar is focused on anthropic inference hooks. So, coming back to hooks, I understand it is helping you expand visibility and control into parts of your organization where, you know, dare I say, the visibility may have lacked before. Can you elaborate on the Inference Hooks use case and, share how you're leveraging Hooks for unmanaged devices? Absolutely. The detector service architecture is great when we can funnel traffic across, right, to to Zscaler. But there are areas… Certain areas where we can't, or there is a path wherein which it will take a long time for us to do the necessary routing or the necessary configuration for us to for us to send traffic across to the guardrails. This is where I think Hooks provides us a great path forward because we don't have to be the proxy in the service architecture follow that particular model, and we can change it in such a way that it hooks up more in the background. Now irrespective of where the user is or the services or the MCP architecture, the hook architecture enables us to get all guardrails across all of our services. And, again, transparently. Right? Transparent not just to the users, transparent also through the services. And we have one place to go without any, you know, issues in between of network traffic didn't route, someone did the wrong misconfiguration, and all of those things. In addition to this, right, when we want to enable… Securely enable AI for our users, we had this caveat of deploying… Had to deploy that proxy based architecture, which is fine for company owned devices, but it does not work for… If that's my personal phone. Right? I still need access to those services. At that point, on unmanaged devices, we can still follow this architecture. It's like the guardrails are always on irrespective of where and what, and we can get the required visibility to secure our business. That's great to hear. Now like I mentioned, we have been working on very innovative solutions, along with your team. So I wanna extend that and tease out an interesting use case. If you could talk about it, related to unauthenticated traffic and other proxies. Yeah. I mean, it's just… It's not always possible for us legacy architecture, you know, as I said, to put a proxy everywhere. We have we have third party proxy services, which we are in in the process of sunsetting, but like anything, but it takes time for us to get back. So we are working closely with Zscaler in order to make sure that this traffic is also redirected fine to the AI guardrail so that we can have the necessary protection for APIs, for MCPs, for everything. For every workload at the end of the day, we need to have the same protection, the same security policy applied across the board so that there is one pane of glass from which we are operating and looking at it, and there are no blind spots. And and this partnership with Zscaler has been great along those lines because we believe in design partnerships, and Zscaler is a great design partner for us. Patrick, I wanna close out this conversation by asking you one last question. What's next for AI in PayPal? What are some of the new business outcomes you are focused on driving next? That's that's a great question. And, you know, that's… I'm pretty excited about what agents bring to the fore… To the forefront and, how we are solving it. So these these things fall back in the same traditional realm of identity and access management, which is which is what I lead here at PayPal. As we look at agents per se, there are there are… And MCP servers and MLMs. There are some fundamental principles that we need to tighten our noose around and make sure that things are secure. And that, to me, that falls in the realm of starting with the traditional identity and access management realms of, okay, who owns this particular agent? What's the governance around it? Is it still required? So providing the agent a strong identity along which it can authenticate itself, but also making sure that the governance pipelines are in place so that agents get attributed right to a human owner. And the purpose of why they exist in the first place is very clear and the intent is clear. That's the first part. Who is the agent? Then there's the second part of it, which is how is the agent authenticating? Is it is it is it authenticating in a way that the credentials are secure? Are they tightly bolted on along with the agent, which should not happen? So how do we externalize authentication and authorization for agent access, agent's access to services? And that should be a centralized mechanism through which we are giving the agent short term credentials that it can then leverage as opposed to as opposed to hardcoded it. So that even if the agent got compromised, there is no exposure. And along those same lines, how are we giving authorization to the agent? The authorization to the agent should only be given for specific tasks based on the intent. This is where zero standing privileges falls into that particular realm so that the agent is only given the authorization when it's required and is revoked based on the intent. And the last part of it is the intent, which is which is more fuzzy than anything else. And I say this because, okay, the intent of the agent is to do read actions at any point in time. But all of a sudden, it's not turned around and trying to delete something. Should that particular thing happen? And if that's not supposed to happen, how do we put the guardrails around it such that those actions cannot be carried out? This is what interests me a lot, honestly, in this particular space that, you know, it's a it's a maturity curve, but we have to reach there faster. And that's my aim and goal at PayPal is to make AI safe, secure, and easy to use for our customers. There are so many threads we can pull there, but we'll end it here. Patrick, thank you so much for joining us. Thank you. It's it's been a pleasure. As you heard, those were some really great insights from Patrick. There's more we are doing to help our customers safely adopt AI across their organization. Now for customers who are interested in getting started with Anthropic Inference and Zscaler AI Guard, It's pretty straightforward. Just a few clicks in two portals, and you are done. Thank you. Scott Bekker Alright. Well, we had a great conversation. Yeah. I had some really positive notes in the, in the, in the chat there just from folks getting real value out of those enterprise, enterprise lessons. So, thanks to, to Ashwin and and Dor and Patrick for that for that conversation. And, before we get to the prize drawing, we do have one quick favor to ask of you guys. So Riley's opening up a short survey that we'd really appreciate if you could all fill out. That's for those of us on the on the actual tech media team who just wanna know, how these sessions are working for you, what you like and don't like. So we'll give you just a minute to, to fill that out. And, you know, what what you say in there will sort of shape the way that that we, do these presentations in the future. While you're filling that out, I'll just mention a few things. Again, there's that, in the docs section, there's that resource link to, the Zoo Scalar page, and it includes some of those some of those things that, that you were seeing there on that that last slide, and, and other great resources. So, be sure to check that out. And we do have one more piece of business before we wrap up. That's the prize giveaway. Today's winner of a $250 Amazon gift card is Pete Benedict from Utah. So congratulations to Pete Benedict. As always, we'll be in touch about claiming your prize after the event. For those of you who asked questions, we will be passing those along, to the to the team, from, from Zscaler. So, hopefully, we can get back to you. And with that, big thank you to, to all of you guys for joining us. It was, great having you here for this actual tech media event. We look forward to seeing you again next time. Hope you have a fantastic rest of

Please provide citations & attributions for any information you generate from this content, thank you.