User Authorization Just Got 10x Harder  Hello everyone. Thank you for joining us. We are just going to take a few minutes as we get folks joining um and then we'll get started. Awesome. Great to see a number of folks joining in. Fantastic. All right, here we go. Um thank you so much for joining us today. Um really really excited about the discussion that we are going to have. So uh this should be a fantastic panel. Uh this lead dev panel is um hosted and created in partnership with the fusion o team. So appreciate the partnership there. The discussion will last about 45 minutes. uh after that myself and the panelists will join you in Slack for more conversations async. So don't hesitate to join and drop some questions there. But while we're live uh I will try to make sure to incorporate your questions. So if you want to submit any questions, please use the Q&A feature in Zoom. Um if you need help, please chat drop a drop a ping in the chat. Um and someone from LeeDev is happy to help you as well. So let's get started. We we'll do some introductions just quickly so you get to know the panelists and experts on this topic and we'll dive into the conversation from there. So to to get going, my name is Dalia Havens. I'm the SVP of engineering at LUSK and I'll be your moderator for today's conversation. So to start off, I'll hand it to you Dan to kick us off. >> Sure. I'm excited to be here. This is a timely and complex topic. Um I'm a senior director for SM strategy and identity standards at Fusion and we're happy to partner with you all today. Uh we're an identity provider like Ozero or Cognto. Um that fancy title means that I spend some time thinking about the future of customer identity access management both at indust industry standards groups like the IETF or uh also looking at the wider market. I also help customers and Fusion community members uh create solutions with our tool and I also love to talk about authentication, authorization, and OOTH. So, I'm I love I'm excited to nerd out about this here. Um I've written a couple of books and I've been in the engineering um practice for about 25 years. Uh looking forward to this conversation. >> Awesome. Thank you, Dan. As ASA, do you want to tell us a bit about yourself? >> Asa. Hi, everyone. My name is Aisha. I'm currently head of application platform security at Aviva. In my career, I had um look worked in security operations, security engineering, cloud security. Uh spent some time at AWS working on security solutions designs and architecture and uh just joined three months ago. uh heading application platform looking after platform security and with that comes a lot of data platforms and of course AI integrations with um complexities. So although IM are not new topics to me um yeah looking at authentication authorization for AI has been a new challenge. >> Very cool. All right to be we'd love to hear about you as well. >> Awesome. Thanks Dalia, Dan Aisha. Really happy to be here on this panel. This is Tani. I lead the customer solutions team for AWS for a large set of customers that run their infrastructure with AWS. And that is why this topic of security overall and authenticating and privacy is very close to my heart because it's priority zero for how we function. I've been in this cloud computing world for over 11 years now. Um and it's always amazing to see how the shifts in technology impact the way customers run their infrastructure and provide us the learning to do better in terms of the infra that we provide. So super excited for this conversation and also to learn from Dan and Aisha from their side of the world how they look at authentication and authorization. >> Awesome. I'm super excited as well. So thank you all for for the introduction. So to kind of frame the the topic of conversation today, we're talking about you user authorization becoming more complex and and that's I think we spent last year sort of in the hype of AI and all the things that that come with it, but it's pretty clear that we are also seeing a lot of fantastic applications from these agents interacting with our systems and becoming more and more autonomous. And that's really what's starting to like add up to the complexity. We are now having to not only think about the humans and what access and level of permissions they have, but now we have to really consider the agents and how to give them the right level of permissions to be autonomous and do things on our behalf. So like the the thing here is that this is not just a readonly access that we're talking about. If we want them to be productive in our workflows and to take actions on our behalf, we do have to consider the specifics and nuances and of course security is a big part of it. When to deprovision some access and all all of the things that we're going to dive into today. Um, so we're going to as as you can tell like this group of panelists are very invested in the topic and I'm super excited to get their thoughts on it. Um, and of course, like I said, I encourage you to submit questions if there is something that you would like to add to the conversation. I'll keep an eye on the Q&A tab. So, let's get going. Uh, before we dive into all things agents, I figured we'd start by framing a little bit the conversation around what we currently have for humans and the authorization models that we have today. and then if you would like to each speak on sort of that baseline and what you're adding in when you think about agents. Um so maybe Aisha do you want to start us off with this question? >> Yeah sure. Um so I I'll start with from my side my my experience like straight away. So traditionally uh like in many other organizations the organizations I worked at we had a mix of uh role based access control group based access. If you really really go back into the past you will see different authentication authorization methods but um yeah not necessarily worried about those anymore. uh when you look at those aback policy base etc etc they were all taken into account a user or a machine and now we have essentially the problem of identity no longer mapping to a person um an agent can act for multiple users um can act for a service um so who who is doing the action is sort of becoming messy now permissions need to be more dynamic. Um, static roles uh may either we will have a all or nothing too too big or too small problem uh with the static roles um and the volumes increasing uh sometimes mid tasks uh the machine escalation may be required etc. So all these complexities uh are bringing uh a lot to scratch my head with my security hat on. Uh so the shift is from how do we authorize a user to authorizing a specific action in a specific context for a specific delegated identity and how do I make that auditable um starting on a negative note but yeah we have now >> that that's that's what we all love to do though. All right Dan do you have any any more you'd like to add? I'd love to hear your thoughts as well. >> Sure. Yeah. And I'm glad that Aisha mentioned Arbback. Um, and I think that it's worthwhile for us to look at some of the other uh models out there because I think she pointed out some good um nuances, but it's good to have a firm foundation. And one thing I always say when I'm starting in these conversations is you're worried about a AI authorization, but how's your human authorization? Right? There's a lot of people who start out and want to jump immediately to the cool new shiny thing, but haven't set the foundation themselves. So, I think you know, arbback, Aback, Reebac, um I'm not going to dig too deep into these pieces, but I think that those are things that you can go look up yourself and uh read up on those authorization models. Um but, you know, setting those aside, you know, the real issue is I think agents get ill to find instructions and they take actions we don't expect. Um, I was talking to a co-orker about this webinar and he said, "Oh, yeah. One time I said to my agent, you know, delete all the files that have the string in it because he needed to do that for some reason and end up deleting an entire folder because that folder had that string in the name." And I think that those kind of unintended consequences where the agent is doing exactly what you told it to do, not what you meant it to do are the reason that we need those dynamic fine grained authorization uh rules that Aisha mentioned. Um people are still learning the capabilities. I am myself and whether you're an end user or developer, security admin. And then I think that our authorization systems are decent at human beings again as long as you implement them implemented them correctly and pretty good at service accounts as well but like this combination of something that's non-deterministic like human beings and moving the speed of software brings in some new challenges. >> Yeah, for sure. Tonvi, I'd love to hear your thoughts on sort of like Dan touched on an example, but I I'm curious what when did you first realize that the traditional authorization models are not going to be quite working for the AI agent world that we're stepping into? >> That's a great question, right? Like um I'll piggy back into what Dan was mentioning about the human aspect of it, right? But before we go there, um I come from a customerf facing background, right? like not that I interact with a product or have a product team but I'm more field-f facing in my role. I capture those inputs to come back and kind of create this space in a better way. So from that aspect I look at this situation in two ways of like authentication is basically a situation where who you are right and authorization is a method to do or prove who you are and the way I realized that as was I was having conversations with our customers they're starting to get more and more worried about these two questions right like how do I know who is coming in and by what method they are coming in right that was my first encounter and clicking as we all started to learn about this transformation of technology from initially traditional AI ML to generative AI to now agentic AI over the course of the past couple of years right so that's where I kind of first um um interacted in these situations the way I kind of have been navigating these conversations is now even more emphasizing the need of a human in the loop as we are starting to use these awesome authentication tools and how we should be able to put some processes and guard rails in place so that we don't have systems that run rogue. So those are kind of the areas I have been focusing on it. >> Yeah, thank you for sharing that. Uh Den, I want to go a bit deeper in sort of like your your your example and and of course like drawing a bit more broadly, but when an AI agent acts on behalf of the user, how do you determine the permissions that it should have? Do you give it full scope of that user permission or like what granular work have you done on sort of like uh the technical details of that implementation? >> Yeah. And I think that you don't give it full scope unless you're like yoloing and you have your entire system backed up and it's not in prod, right? Um so that's an easy guard rail to lay out first. But I mean I think there's a couple of different ways you can think about like setting boundaries around um agents and one is you know be careful with your prompt. Uh if my coworker had said don't delete directories with this string in it but delete files only. Um because I guess in Unix a file is a directory or no a directory is a file anyway. Um neither here or there. Um that can help to some extent that again is relying on the agents um ability to follow instructions. And then um another layer is to kind of build out sub agents so that you have a separation of concerns so that you have an agent that um you know deals with the file system another deals with different uh services. This is less preventing an agent from doing something stupid and more like limiting the blast radius if they do do something stupid. And then the last is really that this deterministic system that we talked about which to be honest with you is still something that we're grappling with defining right like there's some work in the standardization space. Um open ID has an AI community group that they're working to like define some of these things. I know the offzen um which is an open ID standards group just kind of released their 1.0 but uh to be honest with you um I don't have all the answers. Big big shocker. I I hope that the other panelists have all the answers and I can just observe with them. But um I think the fundamental thing to take away and then I'll wrap up is again just reiterating that like this stuff is happening at software speed which means that you need to start to enforce rules and dynamically change rules based on software not uh manual approval processes. >> Yeah ma makes sense. Uh yes I agree with you. We don't all have the answer but like we're we're very much deep into it. Uh Aisha I I'd love to talk about >> if I may if I may intervene piggyback on Dan's stuff right um I think the way Dan you mentioned about having separate agents for um a certain task is super critical as we implement even large scale systems right and now with that even having uh better fault injection threat detection kind of situations is going to be critical because we want to test the systems for situations where like what Dan is saying if a system does anything stupid what's our process to handle it right so that's also critical as we discover answers to how agentic AI is going to transform this uh situation sorry to interrupt >> no no not at all this is actually great because I was segueing very very much into like like let's talk a bit more about like these security concerns and I I think I'd love to hear from you if you want to kind of kick us off with this question but What are the biggest security concerns when it comes to these AI agents having the elevated permission? I think we talked at a high level. Uh, but how are you mitigating some of these risks, especially things like prompt injection or agents being manipulated uh in a way or another? Again, like we're we're talking about like some of these like different than expected results that that maybe we've anticipated before. >> Yeah. Yeah. I want to I want to start with uh saying that everyone is catching up right security is catching up IM is catching up the authentication authorization area even the platforms that offer AI they're catching up as to you know what what's going to happen the use cases are coming up as in when we're learning so we can't necessarily gauge what the use cases will be that will present those risks so we're we're sort of swimming in uncharted territory but there are things we can do So number one and I really want to iterate this and highlight is how well you do authentication and authorization as an organization currently is only going to be amplified with your AI and AI agent use. So if you don't have good practices already in place, you're not necessarily going to bring in this third miraculous solution for your AI authorization that will solve all the problems. So so essential work go and find out housekeeping what can I do better for security to begin with and then we in parallel afterwards we worry about what how else is this impacting our security right. So my biggest concern of course is these are engines that are persuadable. They're they're executing things and we are relying on um our existing authorization methods with them. So that's the biggest concern. Models are getting better in what they can offer um as the model in terms of uh prompt injection attempts being able to identify malicious problems etc. platforms that offer these AI uh uh models are getting better in offering shield, guard rails, content filtering, etc. But what I would like to say is segregation, segregation and segregation. If your back end is segregated enough and you're using your datab your databases are laid out really well, you're really aware of where your data is sitting, you have a good control of your residency labeling and u uh tagging etc. Then um it's going to be easier to put in those multi- aent um and uh fine grain authorization mechanisms. So we're looking at task vote least privileged and look at in your in your input and output controls. Have you got your monitoring and anomaly detection in place with the guard rails logging and auditing etc. So um there is a checklist somewhere out there and I'd be happy to get any questions on on this without going into the rabbit hole. No, this is like very very solid points especially on on sort of like you know jumping to the next complex thing without strengthening the foundation and how critical it is to really sort of look back and make sure that your systems are are in good shape for what you're building on top of as well. Uh Dane, did you have anything you want to add to this question before we transition to the next point? >> Sure. Yeah, a few things. Um, one is, you know, Aisha, I love that you started with the IM excellence being required and I think that that's something we're seeing across all aspects of AI, right? It's an amplifier and I see that with documentation. I see that with coding and it certainly is true with with this this as well. Um, I think it's really important that people experiment at, you know, in controlled environments because they want to build an intuitive sense of this thing too, right? Like even if your CEO is saying, "Hey, we have to do a gentic AI right now. you know in production um let's let it rip you know you don't want to do that right like start with smaller things and there are a lot of people I have a lot of I have some co co-workers who are more skeptical of this than others and find the excited people and like exper you know leverage them and like um you know to experiment with them um as far as segregation I think that's a great point I will say that the only thing I want to add to that is that it is hard to slap a gateway in front right or some kind of external system it a A lot of the fine grain authorization is super business logic specific. It needs to be tied into your application. Um the last last two things I'd say is one is actually I saw a interesting paper that I'll I'll drop in the the Slack afterwards that was actually looking at the prompt that was sent to the agent with an LLM other piece of software and examining it and if the agent asked for uh tools that weren't required to serve the prompt even if it was allowed to with the kind of permission scoping it would deny I access so you can almost use uh the LLM to watch the LLM. Um and last thing is OASP has a an entire Genai project from a developer perspective. That's a a great place to start researching this stuff too. So thank you. >> This is fantastic. Yeah, I'm actually going to like I'm looking forward to the resources that uh you can share in Slack after this uh panel. That that's really cool. All right, Tonvy, I' I'd love to get your thoughts. Like we've talked about like needing to watch and and sort of, you know, auditing and monitoring what these agents are doing with their access. Uh what's different about this with like from logging and observability when it comes to the agent world uh versus the traditional user sessions that we're more familiar with? That's a great question actually and I think the blunt answer to this is earlier uh teams or organizations used to get away probably sometimes with like not having as strictive enforcement rules or going back to Isa's point on um not understanding where the data is so that if anything goes wrong they could do it in a reactive manner as to figuring out who logged in what went wrong the logs and so on and so forth but with AI systems. I think it has become super critical. If you're not aware of how this is happening, it's going to result in situations where it would be very hard hard to revert back um into this into a positive and operational state. So I think that's something of a blunt truth um to keep in mind, right? Um I think um from my organization's perspective and my customers perspective I'm making sure that there are centers of excellence that are established that are um kind of looking at the processes being put in place to verify the agentic AI logic that is being applied to each of these for the lack of better terms microservices that go and execute on behalf of the human right and how are they eventually connecting with each other to create the holistic task that gets executed. Right? So that's um an area um going back to I think Dan mentioned this um this is one more area that I am emphasizing on is each agent should know what their task is and from for that prompting is going to be extremely um critical to make sure that the agents are aware of the environment um and their task right um and then of course real time monitoring guard rails proactively looking at these systems is also so superbly important. I'll leave it at that on this. >> Yeah, that's great. Thanks for sharing. Aisha, did you have any anything you want to add to to this question around like sort of auditing and monitoring? >> Sorry, I could I couldn't unmute myself for a minute. >> No worries. >> Yeah. So, yeah. So, make sure you're logging, you're doing everything that you should be doing in terms of logging. Um so your security uh teams um uh will be will will know what's essential and what foundational login looks like and with any service that you're enabling. So if you're using an example is um um an AWS AI service you're enabling login monitoring around uh around those um and you're if you have um external um identity providers etc that you're using so everything comes together so you can get some sort of correlation as to what malicious could look like. So you're not just storing this log data but you also know how to work with it. um and uh um work with the shields using an additional LLM. I think there has been some latency issues and how is this going to impact my um operations etc that organizations have been asking but some more often than others security is the top priority. So um uh you you just have to consume it and soon absorb that sort of latency that you might bring. But it's important that you know what to enable and then you get your detections from that as well. And there are also external penetration testing, internal penetration, red teaming that you can do to just do some sort of um uh you know chaos engineering testing your the resiliency of the agents that you're using. >> Yeah, makes sense. Yeah, go ahead. Before I just I wanted to ask uh Tommy mentioned uh centers of excellent excellence and I just want to know kind of like what were the goals and the outputs of centers of excellence um what she's found useful uh in terms of that. >> Thanks Dan. I think the centers of excellence are having um most importantly business goal alignment right and stakeholder alignment they need to and I think you mentioned this point around making sure that hey if there is a top- down push on just like going and creating agents and agentic AI there should be a process laid out uh from a business and strategy point of view as to how we are going to do it what are the systems we're going to start with um what's our plan B in case things goes go wrong right and then how do we align that to the broader organization ation's vision to prove the value proposition of the agentic AI systems that we are putting in place. That's the business side of it. From a tech side of it, I think the way I'm looking at agent or encouraging the coes to look at agentic AI is what actually is your biggest pain point. Right? Don't put in agents for the sake of putting in agents and looking cool about it. Right? You need to address what your biggest pain point is. be it for your operations teams, your IT teams, your dev teams, your QA teams and so on and so forth. And then prioritize by asking a couple of questions and surveying a smaller set of people to create like a a prototype kind of a a function that then you can test in a sandbox environment and then ship out to a larger uh production environment. And the best part with agents is there is availability to test across multiple agentic AI tools and functions to make sure it suits and validates the checkpoints that they that we have put along the way. >> Yeah, thank you Tommy. This is this is great and Dan I really actually appreciate it. Uh so like with this why don't we transition into talking about the organization? So, so I'm curious um about like all of you feel free to jump in, but maybe I'll direct the question to for Dan to start us off with, but when it comes to like who owns agent authorization decisions in your organization, so like we talked about the technical side of things and how we we might consider our implementation and access controls and all of that, but the organization plays quite a bit of a role in the future of the world that we're operating And so just curious if you want to start us off with this question and your thoughts on it. >> Sure. And I will say that like I think that an important place to begin is the size of your organization, right? And this may shock some of you, but Fusion Op is smaller than AWS. Um and so we have a kind of a different different kind of approach probably than AWS has, but uh we have a combination between engineering and security and depending on the resource even the IT team gets to kind of weigh in. Um, we're in the early days, just to be totally blunt, but I think one foundational principle that we've all arrived at that makes a lot of sense is that the human is still responsible. So, if you're having agent build out code or build out content or build out, you know, handle business processes, do financial planning or whatever else that agents can help with. At the end of the day, human being is responsible for managing um for owning those outputs. Um, and that's I I I understand the beauty and desire for autonomous agents that just kind of go off and do things that you want them to do, but at the end of the day, there's a human being who's kicking off that process, who has that desire, and to tie things back to that human being, to that human identity is really important. >> Yeah, makes sense. Um Aisha when it comes to like compliance so like from from that perspective your sock to your GDPR and so on how do you handle the fact that AI agents are accessing and and potentially modifying user data? Uh so what does consent even mean in that context? If you want to you know take us through it and elaborate on that part. Yeah, obviously with um um data access and u human or non-human authorization be coming into the mix with AI agents and we wanted to be able to do a lot. So the the stuff that I explained in the beginning basically presenting um a picture where we're now blurring the lines between what it is accessing so the scope is widening. So if you're looking at scope to type type two and then you've already scoped it and now there's an AI and that's touching up on a whole lot of other things then your your audit will need to include those elements environments etc within within your scope right so that brings those challenges um in in addition to that uh it brings the challenge of uh being not only the audits and auditors coming in um armed uh understanding it fully and being able to uh not only articulate what's being asked but also understand that it is in place because those are all up in there to be very honest with you. Um we'll we'll find some confusing or confused language um in all this I haven't got to uh those yet. There is some definitely a huge learning curve for myself to understand what that means and yeah so the boundaries being blurry what does this mean for GDPR and uh granular consent it's not just the yes or no it's scoped permission you know are you going to do immediate revocation um maybe you're going to need another AI agent for that >> where they have the permission um >> yeah uh risk assessment again becomes uh a challenge with the processes the sensitive data and is it scale and then you need to document those risks and mitigations as part of the audits as well. >> Yeah, makes sense and a quick question. >> Yeah, go for Yeah, I just want to jump in and I mean so I read this, you know, I thought about this question. I thought about, you know, um what you're saying Asia and I guess my question is how is the GDPR compliance story for AI agents any different than it would be for a normal like a software system that's accessing PII? Like is it is it you know a difference in kind or a difference in you know scale or or is it just kind of the same rules apply? we just it's fuzzier because the it's also new. >> Um it's it's also fuzzier because uh there's not necessarily a a a clearly defined back end. Uh it's quite dynamic. Um it keeps moving. Um you want it to to do the best it can do and it can do the best with data. Okay. So if you scoped the segregated gap data based on your regulatory requirements then you're limiting your model from access. Um so there's that sort of conundrum as well. Um, I guess it's not necessarily any anything different if you know exactly what that AI agent is going to do and you got you got a clear use case and that clear use case is going to uh provide that and it's just a matter of you know complying with GDPR as you would if it was a new type of software right yeah I guess the complexity comes in when um you got uh a business case that is to enable AI rather than a a complete clear problem that you're using AI before is when it becomes really >> right which I love that goes back to your point about the guard rail like having a firm understanding authorization model at the beginning and Ton Tony's point about like got to drive a business case like don't chase the shiny object do it for business reasons and you're going to have a lot you know it's going to make it easier to have success criteria and easier to comply with the law who knew that's great >> love it uh yeah there was a question from Andy uh very relevant to to this particular conversation that we're having. So, uh if any of you would like to go into the Q&A and elaborate, but I think we've touched on uh the answer to it um you know with with the points that you made. So, fantastic. All right. Maybe we transitioned a little bit future forward um and start maybe like talking a little bit about you you're deep into this world. Uh what are your predictions? What are what are some things that that you feel will look very different or have we sort of converged and you like to the conversation before the human like whether it's human or agent it's still around the same problem definition that we really need to think about. So I'm just curious um if any of you want to share your future forward you know ideas with the with the group here. I mean, I guess I'll go first. I mean, I think that that, you know, we are going to move towards a more unified system. Um, I think if there's one thing I've learned over 25 years of software development is that people like convenience and agents are driving convenience, right? Like they're making things more efficient. They're, you know, it's a lot easier. I do this myself. it's a lot easier for me to ask um an LLM or a tool to give me a summarization of a meeting or to keep track of a meeting so that I can actually sorry keep track of the to-dos in a meeting so that I can actually pay attention and like contribute than it is for me to take notes. Um, so I think that this wave is coming. Um, and you're going to start to see human beings, agenic systems, and don't forget those deterministic software programs which still have a place all have a single source of identity. And then I already kind of mentioned that paper and I think keeping an eye on um the research that's happening because where things were a year ago is not where things are now. this is less of a prediction and more of like a warning. Like you can't just have a policy and and be kind of one and done. Like things are changing way too quick to do that. >> Maybe I can add Dan to what you were saying, right? I think I'll first go into predictions and then warnings. Um going back to your point, my predictions is around this is going to be a part of day-to-day life. Even if we try to be skeptical or try to avoid it, I doubt we're going to be able to do that. So that's another hard realistic um presentation. There are a couple of good podcasts that I listen to that I can even add in the Slack channels as we follow up that gives a perspective on this across industries, right? Um I think the other prediction is some of the low-level lifting that all of us in every job family had to do to accelerate uh operational excellence, productivity. I think that is where we it will do the lowhanging fruit stuff in a pretty systematic way depending on how we configure it to do it. Um I think that's the second prediction that would get easier and the warning is um I think we all need to start getting handson and dirty in terms of playing around with these systems and learning and capturing feedback. uh because the more we are disconnected from this world I think it is going to get harder and harder to kind of make sense of where the business and the organizations are headed and also to leave a mark right I I've started comparing our tech roles to being doctors in the future is like hey anytime any new surgery new medical invention comes in the doctors always needed to be specialists or be like deep into the inventions I think this is the era for people in tech to be specialists and surgeons and start um experimenting and learning through hands-on work. >> And Aisha, if you have predictions, I I I would love to hear them, but I did want to ask and maybe this is something to do after Aisha gives predictions is uh any surprising results from experiments because I agree with you, Tom. Like getting hands-on is the only way to really understand. I have a something that was surprising to me that I'd be happy to share, but like I don't want to take away from Aisha's predictions because I want to hear that too. So, >> no, I'm a lot more curious about what you're going to say than you know what I was going to say. Like to me, it's just going to be very very faint. No, so I'll keep it short because I want to hear what you're going to say. So, yeah, I I pretty I think the same as you you uh Dan Tan, right? Um I think we will see uh we will settle a little bit in what we uh understand that we can expect from AI to give us and um and how as we understand more and more uh we will also have better um we would better equipped um to get it to do things in the way that it's not going to damage us and provide uh problems. So yes, you use a authorization um could be harder, but it also means it can be easier if you're doing your data authorization in the background a little bit more cleverly and how you're working with your data. I I read a an AWS blog post on this um uh previously and uh it's it's so I'll I'll post a link um it's it's quite interesting giving it a different way of thinking. So as things are becoming more data, you know, it's going to be more data in my opinion, more data focused from authorization perspective as to what happens with AI agents and how how much of our lives will be of it, consumed by it, included in it, interacting with it. It's uh the trajectory is definitely up. Uh in my opinion, it'll be interesting. Good times to be alive. >> Yeah. All right, Dan, are you gonna share with us? Uh, don't leave us hanging. >> I feel like I met I built it up too much because when I first encountered AI, I mean, I did some stuff even in pre-hat GPT, but like I had to like create poems, right? Or read a blog post that was like and it was totally mechanical blog post. And the things that really have surprised me, two experiments I've done are one connecting to my task tracking system and just saying, "What did I do last week? Please summarize it for me." so I could report that up. And the second was um I had a bunch of uh files that were written by people who are non English speakers and I just said please correct the grammar of and the spelling of all of this stuff right and it wasn't so surprising in that it could do it. um it was surprising how good it was at doing it and it even corrected some things where it was like this is an incorrect link right so um and those are small things and I think that this goes back to mying point is like don't start out with the multi-genic system start out and learn the capabilities of these things because we do need to as um said get our hands dirty so sorry if I blew that up too much I'd love to hear what other people had that was uh the interactions with AI that they were surprised That that's really great. Uh there's one question that that I want to see if um if we want to touch on before we wrap up that that came through the Q&A. Um so uh the question goes, I love the idea of LLM's monitoring LLM. Curious what patterns or architectures are teams using to secure that monitoring layer itself? >> Oh, go ahead. both Dan because I think you post mentioned it and I'll add >> I mean I mentioned as like an experimental thing that you got to keep an eye on. I did not mention it as like a real full running system that I could answer that question with. So and uh the one thing I will say is that paper I did mention wasn't just LLM monitoring LLM. It was also deterministic systems monitoring LLM too. And so don't sleep on the fact that you can do things with deterministic systems as well. Yeah. >> Yes. Go for it. Sorry. Go ahead. Go ahead. Sorry. >> Yeah. There there are a couple of ways. So you can do live session integrated in as AI agent is interacting with the user in the use case of let's say a chatbot then you can have an additional uh model that will be in between. We'll be looking at in addition to the models own prompt injection um shield etc guard rail that they have natively or the platform they're using you could use another third party depending on how much of a restriction you want to bring in right that's that's one aspect another one is um when I was so previously I worked at AWS when I was there I did a a a talk I think I'll give a link as well I did a talk of the the three principles of AI is one is AI security of AI and then AI for security and security against AI um because we're now looking at AI threat AI stem threats coming in so how do we secure so when you I I'll send a link if I can find so when you look at that there will be other options where you can integrate in AI I based solutions into your overall security controls and capabilities. >> That that would be great. Yeah, I I'm really excited actually about the follow-ups from this panel and all the resources that that each of you might want to share there. So, I encourage you to like I encourage like the panelists and the participants to join the Slack channel to continue this conversation. This was clear like really fantastic. I really appreciate the the notes and experiences that you're bringing to the table as we've talked today like the the world ahead of us does not leave the fundamentals of what we we've learned about the humans permissions and authorization. So like that's still very very important and as we layer in agents and continue to deploy them in very useful use cases which will allow them to take actions on our behalf. we need to continue to augment the systems that we we have to sort of get the problem solve and I think Tonvi I really like that framing of like it's not just like AI for the sake of it but we are truly like I I would say like for this year I'm super excited to start to get into the value of the application uh itself and and what more complex problems we can possibly solve that maybe weren't as attainable before. So there's a lot more to look forward to. Uh we are going to wrap up here since uh we are at time but I want to thank you again for for joining. I want to thank the panelists for preparing ahead of this and joining the conversation. This was really fantastic and I very very much appreciate your time and and all the that you had to share today. So thank you for that. All right everyone. >> Thank you. >> Thanks. >> Great to see you. Bye everyone. Talk to you later. See you on Slack., User Authorization Just Got 10x Harder, YouTube, 01kg84nc60q236nwkfhwj7r2c0.