Sponsored by: OneTrust | Operationalizing the DASF: Reducing AI Risk with Runtime Enforcement

Summary

The Databricks AI Security Framework (DASF) can be extended with policy-driven runtime controls that enforce consistent access, use, and governance of data across analytics and AI workflows at enterprise scale.
OneTrust and Databricks combine risk intelligence with data and AI infrastructure to turn static governance policies into enforceable, auditable controls that operate inside the platform in real time.
As AI agents proliferate across enterprise workloads and insurers begin excluding AI-related damages from corporate liability policies, scalable governance is becoming a competitive differentiator rather than a compliance obligation.

Chapters

FAQs

What is the Databricks AI Security Framework (DASF)?

The DASF is a framework for operationalizing trusted data and AI governance on the Databricks Data and AI platform, providing a structured set of controls that organizations can apply across the AI lifecycle. This video shows how it can be extended with OneTrust's policy-driven control catalog to enforce consistent access, use, and governance guardrails at runtime.

What is Unity AI Gateway and how does it block risky agent actions?

Unity AI Gateway is a technical control layer within Databricks that can detect and block high-risk agent actions before they reach production systems. In the demo shown in this video, it applies automated remediation protocols when AI violations are detected, providing enforceable governance rather than advisory policies.

Why is human oversight insufficient for governing AI agents at scale?

As AI agents proliferate across enterprise workloads, the volume and speed of autonomous actions quickly exceeds what human reviewers can monitor manually. This video argues that governance must operate at machine speed, using automated detection, blocking, and remediation to scale alongside AI ambitions.

How does the OneTrust and Databricks partnership approach AI risk?

OneTrust brings risk intelligence and a DASF control catalog to the Databricks Data and AI platform, while Databricks provides the infrastructure where governance is enforced directly. Together they offer an AI risk self-assessment tool and maturity benchmarking capability, allowing organizations to identify gaps and demonstrate compliance without disrupting innovation.

Full transcript

[00:07] All right, let's go ahead and get started. Everybody appreciate everybody taking the time to attend today's session. I'm Shane Wiggins. Uh I lead our AI and data governance at One Trust. We serve more than half of the Fortune
[00:20] 500 and today honored to share the stage with Praep um senior staff partner engineer with data bricks. So over the past couple years, One Trust and Data Bricks have grown this partnership and
[00:33] we're really excited about some of the latest announcements that I'll take you through later today. One trust brings the risk intelligence. Data bricks brings the data and AI infrastructure and then together we put governance
[00:46] where the work actually happens inside the platform in real time. So a quick view on the agenda of what we'll cover today. I'll cover the first section, Purdep will cover the second section and then I'll take each one of you um through a demo towards the end. So AI is
[01:01] poised to define the next decade in business, but we believe the the real differentiator won't be just who adopts AI, but really who's going to to govern it well. And fundamentally, we believe
[01:13] that the companies that are going to get the most value out of AI are those that are both innovative but also intentional about governance. Uh just last week I was on site with one of our largest uh
[01:26] insurers, global insurer, and their CIO asked me, can our governance model scale at the same rate as our AI ambition? So we'll play that back. Can our governance
[01:39] model scale at the same rate of our AI ambition? And what we're seeing is a lot of leaders are starting to ask this the same question. And part of that is a result of governance getting amplified
[01:52] through a couple different forces that are starting to converge uh together. So first you have agents starting to peripholate across your your business and your enterprise workloads and applications. You have Genai and agents
[02:05] that are making these workloads both stateful but also non-deterministic. You have regulations both globally here and in various states as well that are starting to force direct accountability
[02:18] and also enforce significant penalties. And the other piece that's interesting is we're seeing insurers. So you have Birkshshire Hathaway, CHUB Travelers and they all have filed to exclude AI
[02:32] related damages from corporate general liability practices and policies. So what this means is essentially the message to the CEO is you're on your own. And then finally, what we're seeing is customers are starting to require
[02:44] more transparency as a condition of doing business. So two examples is, you know, I sit within R&D, but I can't tell you how many sales conversations I've been in the past six months to help our
[02:56] sales team actually transact our product, help the field understand what AI we're using within our product and services, what underlying models are we using, the capabilities, and ultimately how did we govern that? And the second piece was talking to the AWS team uh
[03:08] yesterday and what they're starting to see is these governance and transparency requirements are starting to become available within the RFPs and this is something that they really rarely had encountered in the past as well. So
[03:22] between these you you have these different forcing functions which is really making governance absolutely critical for the business. Now we took this and validated this uh in the field. So we surveyed over 1,200 uh IT security
[03:34] compliance and AI professionals and the most telling statistic coming out of that was that 70% say their ability to govern AI is often at odds with the
[03:46] speed at which AI initiatives move. So this goes back to the same conversation with the CIO last week saying that you know governance cannot keep up with the pace of innovation. And if you think about this it makes sense. In the past,
[03:58] governance was built around periodic human-driven processes. You could think quarterly reviews, waterfall type developments, and well-defined applications that were for the most part
[04:10] static. But AI doesn't operate in that rhythm anymore. These systems are now running continuously, are adapting in real time, and increasingly taking actions without human intervention.
[04:22] So if we bring this back to, you know, looking at the pace of innovation, you take that plus the overall democrization of AI, what used to require specialized teams with access to advanced hardware
[04:34] can now be deployed to the cloud in just a few clicks. So that the tech is moving faster than any review board or AI governance committee can actually meet. In the past, you would have your risk evaluation and you would go through an
[04:47] approval cycle where you would have security, privacy, legal, and ultimately sign off. And by the time you get an answer and if you're lucky with an approval, the model is already versions behind and the project is already in
[05:01] production. So governance cannot run this core behind the thing it governs. the answer has to land in seconds on the projects that are already in flight and the AI that's already in use. And this
[05:14] is really fundamentally the challenge that we're seeing organizations face today. Balancing the speed of AI innovation with the need for effective oversight. Uh on top of that,
[05:26] now if we bring that to first principles, agents can scale exponentially. human oversight cannot. So you can continue to add reviewers, you can stand
[05:39] up committee processes, you can stand up these approval steps, but those are going to scale linearly where the number of models, agents, and AI powered workflows are going to continue to multiply. So this means the answer isn't
[05:53] just hiring more reviewers or adding more checkpoints because no organization can outstaff this exponential growth curve. But what's needed is a different
[06:05] operating model. And this is one where governance is embedded directly into the AI life cycle with policies enforced programmatically
[06:17] continuously, but also at machine speed. And governance must scale with the agents themselves, providing these guardrails in real time rather than trying to just catch up after the fact.
[06:30] Now, everybody's obsessed with making AI smarter, but you know, we all know that the smartest AI doesn't always win. Now, we believe that the the AI that's going to create the most value will win as a
[06:43] result of a couple things that you need to get absolutely right. And we call this the agent success quotient. So, you have IQ, SQ, and GQ. And we look it as a quotient for a reason because if any of
[06:55] them are zero then the sum of the outcome is zero as well. So the first one is the intelligence quotient. This is the obvious one. Does the agent actually work? Can it answer it correctly? Can it complete task
[07:08] reliably? Can it achieve the business outcomes? Now this comes with better prompts, better workflows which brings us to the second one around security. Now, this is critical to ensure that the
[07:20] agent can only access what it's supposed to and is protected from outside threats. And as part of this, you really want to ensure that sensitive data is protected because an agent that is intelligent but ultimately insecure is
[07:33] really no longer an asset. It quickly becomes a liability. And then the third factor that many organizations underestimate is GQ. and not the GQ that Purdep is featured in, but the
[07:47] governance quotient. And this is where we're seeing a lot of organizations struggle. And it's not necessarily about whether the agents are just working. It's about whether they're allowed to work. Does it comply with regulations?
[07:59] Does it respect customer consent? Is it using data appropriately? And this has really been the focus of One Trust the past couple years, which we'll highlight um later on in the presentation. The other part of this is the the human
[08:12] tension and the oftentimes organizational conflict that sits underneath this. So oftent times in these review processes, you'll have two different groups. You're going to have your AI owners who are shipping the
[08:25] initiatives and ultimately delivering the ROI back to the business. But then you also have the AI governors and these are the folks who help to define the policy and also really protect the the
[08:37] business. And what we're seeing too often is they'll start to collide in the middle. So there really isn't a real bridge between them. It's much more of a fragmented handoff and that quickly
[08:51] breaks uh under speed. So done right, the governors can set the rules once and then the owners move faster because of them because of governance almost being invis invisible to them in the process.
[09:04] Now all this has gotten more complex as AI proliferates. So Gartner frames four deployment classes um that organizations need to consider when they think about AI governance. So on the buy side you
[09:18] have AI in the wild and these are going to be your public tools that people already use either with or without permission. Now the embedded AI these are the features that are riding inside
[09:30] the SAS that you already pay for. Now if you move to the right you get more into the build side. So hybrid AI is where you start to build out your agents assembling models datas and the in the various agent platforms that you're
[09:42] working with into something that's unique uh to the business. And then the final one is AI and house. So these are the models that you train run yourself. Think about more traditional and predictive ML. What's interesting about this is you're
[09:55] starting to see a blend between embedded AI and hybrid AI where a lot of the current SAS that you all procure day in and day out, they're starting to extend it into their own agent platforms which allows you to then build the agents as
[10:08] well and you start to get this this middle ground area which is very difficult uh to govern at scale. So how do you get started? Uh we built a framework to really guide our customers through this journey. It's been
[10:20] extremely wellreceived uh this past year and it really starts by establishing the right foundations that are going to allow your teams to to scale. Now, every organization sits somewhere on this curve and where you land really
[10:34] determines whether governance is going to be a tax on innovation or end up being a force multiplier uh for your organization. So, let's go through these. So at stage zero, customers and organizations are
[10:48] mostly reacting to regulatory pressure with uh very fragmented efforts. The projects tend to not be centrally tracked. Documentation is manual and risk assessments tend to be inconsistent
[11:01] and at this stage governance tends to feel like a side project and it's often owned by privacy or compliance teams really without a dedicated budget uh or resources. Now at stage one, organizations are
[11:14] going to start to formalize governance. They are cataloging projects and starting to build out these compliance review workflows, but it's still quite slow and manual and regulatory alignment
[11:26] tends to be one of the main drivers. Teams want to fast, but oftent times you'll have over reliance on these manual checks and again this is going to create the bottlenecks and ultimately a lot of the friction uh within the
[11:38] business. So what do they do from here? they tend to explore integrations and look at how we can automate better. Now stage two is going to be the turning point. This is where governance becomes embedded in the AI life cycle. You have
[11:50] policies that can now be translated into automated controls and you have these checkpoints that are actually going to align with the development milestones and the oversight no longer slows the
[12:02] business down. So here you have compliance and your technical teams actually starting to work together. And this is starting to enable uh faster but also safer AI deployments. And then stage three is the most mature
[12:14] stage. This is where governance quickly becomes a competitive advantage and where we seeing a lot of our customers start to win. The oversight now scales across thousands of models and agents with programmatic enforcements and
[12:27] runtime guardrails. And organizations here can now demonstrate responsible AI to regulators, to their customers, to their boards, and still continue to innovate at speed while still
[12:39] maintaining trust and accountability. So the way to read this is that most customers today are still on that left-hand side of the curve. They're in this activation stage. They're starting to stand up inventories. They're
[12:53] launching these intakes. They're building their policies. They're assigning risk. and they're really trying to operationalize their AI governance program. That's fantastic. That is work that is necessary, but the real unlock is helping customers cross that chasm into activating AI governance
[13:07] to really scaling AI governance. And I think we see that chasm as a tremendous opportunity for each one of you and where AI governance starts to become part of the enterprise AI stack.
[13:19] All right. So if you're going to take one thing away from today and what we're going to uh highlight uh in the demonstration, it's the importance of translating AI risk into enforceable
[13:31] controls. So again, how do you translate AI risk into enforceable controls because every runtime event, every prompt is now becoming a policy decision. Now if you look at uh the the
[13:45] diagram here on the lefth hand side customers will have their risk and compliance requirements. This could include privacy, this could be data protection, this could be information security, this could be your definition
[13:57] of AI acceptable use, could also be thirdparty data agreements, regulatory obligations and then certainly the frameworks from our folks here at data bricks uh as well. Now historically customers have tracked these you know
[14:09] requirements through documents, assessments, policies and approvals but in AI that's not enough. Those risks and these associated controls need to become machine readable rules and get activated
[14:22] where AI and data get processed but also enforced at both the design time and also at runtime. And you bring all this together you have tremendous observability which helps to close the loop. you get the runtime signals that
[14:35] are really telling us whether these controls are working uh or not. So this is a fundamental shift for moving from governance as documentation to governance as action and allows you to
[14:47] ensure that every documented risk is traceable to a live measurable enforced control across every cloud every model and every agent. So now what I want to do is bring up Pradep to help take us
[15:00] through the data bricks AI security framework which is a tremendous input to help go from these controls uh into enforceable controls.
[15:19] All right, thanks Shane. Uh all right, am I audible? All right. Um so you just heard the problem, right? Um risk reviews basically take weeks. um you know sometimes months uh you have to translate all the documents to you know
[15:31] controls and apply them actually make it happen technically it's it's a very challenging problem generally and uh now that you can see agents everywhere every platform every vendor every every
[15:43] everyone is building agents so uh it's the problem has become much bigger and we need a scalable solution essentially right um good thing is we we have something written down for you you don't
[15:55] need to figure it out all all on your own. Um that's what data bricks uh AI security framework is. Um I'm going to quickly cover about you know what this framework actually means and uh we're
[16:08] going to walk through we're going to unpack it a little bit. Uh you know there are all these different terms you know I want to unpack it and then just walk through a risk end to end. You know how the how does all this work? how is
[16:20] it listed on the framework itself and how how you can take advantage of it and how one trust works with it essentially. Um and honestly if you want to take a photo of one slide it it it should be
[16:32] this slide uh you know you got a QR code there uh but it's gone now we I put the QR code here as well so you can uh you can take a picture of that. So um we
[16:44] actually released a new version. I mean the photograph in the web page says version 2.0 but it's actually 3.0 so you can check it out. So what is this? It's a it's a framework. Uh it's free um and
[16:56] uh we shipped a new version this March. Um think of it like a checklist. It's not like a whole textbook. Uh so it's it's more of a checklist. Uh and it breaks down um essentially your AI
[17:08] system. Of course now it's it's not just a system. It's an AI system. Uh there are many parts to it. So uh it lists 13 parts. It breaks it down into 13 parts. Um, and then things like your data, your
[17:22] models, your promps, your agents, all those things are parts. And then basically for each of the part it will say what could go wrong in that part, right? And uh then it says you know those are the risks essentially what
[17:35] could go wrong. So it lists out like 97 risks essentially. I'm sure those things those numbers are going to climb up uh as they you know as we identify new ways people are you know facing risks. um things like it could be things like data
[17:48] poisoning for your training data uh it could be things like prompt injection most likely you have heard about that I'll walk through that as well um and then uh things like an agent leaking uh some data your sensitive data through a
[18:01] tool like you know all these MCP tools you're connecting through and then you're giving agent access to that it can just like hey you know what yeah let me just get some sensitive information send it back to the tool and we don't know where that what that tool is doing
[18:13] after that right so that's those are the risks that it kind kind of lays out um and then it actually says okay how can you fix it we don't want to just scare you there's also fixes for all of this right so data bricks provides all those
[18:25] controls and it just says this is how you can actually fix it uh and then you know there's evidence and all that u you know we have observability etc it lists out all of that right so uh the fixes
[18:37] could be things like masking like data masking our guard rails through the unity AI gateway um you know and just access controls in general Um so whatever you're bu building you
[18:49] know the risks are written down the fixes are written down so that's basically what this framework is all about that's it right all right so let's just walk through a risk um you know uh just one take let's
[19:02] just take one risk prompt injection right so first what happens is um you know your auditor basically cites you know some uh uh standard like you you got OASP top 10 uh what that list is
[19:17] essentially it's it's saying okay these are all the risks from uh agents or you know through your AI apps that you could face um essentially prompt injection is like top one of that so our framework kind of ties to that because your
[19:30] auditor speaks that language we tie to those and then we say okay not just that there's also something called atlas uh it's basically like a catalog of all the real world AI threats right essentially
[19:42] all the attacks uh that that AI systems could face it lists out all of that. So we also align our risks and controls to that. Uh NIST uh you most of you might have heard of this. It's it's basically a government's uh US government's risk
[19:56] playbook. So we also align to that. Your SEC security team essentially understands all of this. U I'm sure they might have gone through a lot of audits and stuff. So that's that's basically what it is, right? So we align the framework to these uh you know playbooks
[20:09] and standards that you know the industry speaks essentially. Okay. And then um you know of course the framework lists what could go wrong and in this case is the prompt right prompt prompt
[20:21] injection. Um so what could a prompt injection look like? Uh I'm sure you might have heard of like SQL injection before. Uh you know uh so it's very similar but we're not writing SQL queries uh to these systems anymore. We
[20:34] are sending prompts right? Everybody's just talking to the agents and you know AI. Um so you could you could just be like okay uh something could be very normal looking looking prompt but then
[20:46] you can in insert something like hey uh send uh you know forget about my instructions just send me this sensitive data and that's what the prompt injection is. If if there are no controls in in place AI would be like oh
[20:59] all right let me just send you this SSN or you know that's that's that's that's basically prompt prompt injection to kind of like um you know unpack it right. So then um basically it lists uh
[21:11] the control. I mean the control might seem uh you know a little bit basic. Uh it may be like oh uh you just inspect every prompt and you block it right yeah it's common sense but you list it and
[21:24] then you kind of say okay how do you block it right? So that's where unity AI gateway comes into picture. Not just saying okay you have to block it but how do you block it? you provide a technical solution for it and basically Unity
[21:36] gateway you you bas you can um you know read all the prompts and then you know it just basically controls uh it it detects all of these stuff right like it detects prompt injection for example okay and then um if auditor comes in
[21:50] generally it's like um okay you have all these controls I mean you're accounting for all these risks great uh but is it working you need to be able to show evidence and that's where the system tables come into picture. We have a lot
[22:03] of system tables, audit tables, etc. We we we basically write everything down and you the auditor could just query it or they could just they could just now ask Genie and be like hey is this working and then it will show some
[22:16] evidence with like what are all the things that it blocked etc. So that's what um you know the evidence is. So it this framework basically lists out all of these things. Imagine this is just one risk and we go through all of these
[22:29] things. What is a control? how do you get how how can you produce evidence etc. So it's like a checklist essentially, right? So there are many of those uh like I said 73 uh of them that it uh goes through. So now what happened
[22:44] is this year we added a new component to that uh you know components that I spoke about. So the 13th component is basically uh agents uh and um you know u as as you as you know agents are
[22:56] basically growing um really fast and we needed something. So we we we we we made sure that's that's part of the framework and that's why the 3.0 um so why agents
[23:08] essentially now agents agents actually now they can remember things they can actually use tools they can talk to each other right they can also take actions that's the worst part actually uh so so
[23:21] in 3.0 So we what what we did did was we included this and then essentially there are like six six uh controls and 35 risks that we lay out. uh for example one of the risks is they remember like
[23:35] like I said memory could be poisoned right uh that's what this means and then uh what is the control you have to keep human in the loop right so and then they can use tools right um you know a bad tool can basically hijack u a system
[23:49] right so for that you need to give least privilege so it can't really take those type of uh dis um actions dangerous actions and they can also talk to each other So if one agent is corrupt, they
[24:03] can just be like okay let me just across all agents. Um yeah so you you basically trace every step you see what these agents are actually doing uh and that traceability is available to you that's
[24:16] a control and then they can act like they can take action like I said right so uh we call it the lethal trifactor so essentially a rage agent that can that can read your sensitive data uh and then
[24:28] it can take input from an external user and then it can take action right that's the lethal trifactor so for something like that you want to put guardrail in front and that's the control essentially.
[24:40] Yeah. So to summarize right uh this framework is basically a paper right uh it's not going to stop an attack in the morning uh or when you're sleeping right um so you it basically lists out all the
[24:54] risks controls and then hey you can produce evidence etc. Um and all of these controls are uh you know they're they're not they're they're either the UI interfa like you have the user interface to turn on these
[25:07] configurations etc or apply these controls or they speak SQL and uh APIs right so you have to do it still so that's where one trust come comes into the picture and they fill the gap uh they actually convert all of these
[25:19] things all of these all of the framework the controls and the risks and all that into u actual uh technical configurations on data bricks automatically Right. So, u Yeah. Anyway, uh so Sean is
[25:32] going to show you guys all of you know how it's going to work. Uh so, yeah, thank you.
[25:44] All right, thanks Praep. Now the fun part we get to actually show you the product. Uh but before that do that just wanted to you know remind everybody the the progression here. So as Praep mentioned we start with these policy requirements and we have the risk
[25:56] obligations. We then translate them into machine readable controls and then actually enforce those controls uh directly within the the data and AI stack. So combined really the goal here is to take governance and the intent
[26:08] that comes from your compliance and risk teams and actually turn that into something that can be applied consistently automatically and uh at scale. All right. So identifying uh a data risk
[26:23] means nothing if you cannot uh really act on it. So, one trust maps the regulatory frameworks like the the EU AI act into required controls. Now, we took that same approach one step further with
[26:36] data bricks. So, together we seated the data bricks AI security and governance framework and this being that prescriptive set of controls that Pere took us through really built for data AI workloads uh running within within data
[26:50] bricks. So this is the framework that's you know live within one trust you have every control ready to um operationalize and if you start to drill into any of them so we could take data classation
[27:03] and quality now the description the guidance the framework mapping it's all in one place so now you know exactly what good looks like and this control is also mapped to the policies procedures and standards that ultimately enforce it
[27:16] and this is really the jump from knowing the control to actually uh running it. So let's go and jump into to data bricks. So here's a thirdparty smart contract vendor and what it's doing is it's accessing utility uh customer data.
[27:29] So it's going to be through a data bricks workspace and as you know this pipeline starts to expose customer records our platform is going to continuously scan the data and actually identify these categories of data such as personal data, payment details or
[27:42] even energy usage patterns. Now uh the vendor uh query request is going to go and request full customer records. Now what we're going to see here is this includes fields that should really never leave the the govern environment. Now as
[27:56] a result of this one trust is going to automatically detect this as a policy violation. Personal identifiers that should not be um exposed in these agent interactions. So rather than generating a ticket for a human to go and resolve,
[28:09] that policy definition is now going to be directly uh absorbed into the agent and MCP protocol layer, which can now allow you to actually govern the data that the agent can access and the
[28:21] specific tools that it can call. So in this case, I'm going to take a specific field and go ahead and actually enact that control. And what this means is that these masking rules and the access restrictions um aren't necessarily
[28:33] applied after the fact. they're now enforced at the point where the agent's going to make these type of requests. Now, these aren't just generic API calls as well, bolted on from the the outside. So, the masking and the filtering
[28:45] controls, these are built to data bricks' own architecture and actually certified by data bricks as well, embedded directly into the platform's um governance uh layer. So, you can see here now that I go back and run that same exact query, I now have that
[28:58] control automatically applied as part of that agent interaction. Now in addition to the the masking and the sensitive categories of data, we can uh also automatically detect additional policy violations for records where customers have not consented to their
[29:12] data sharing with this vendor for AI use. So now you have records from non-conenting customers which are um going to be excluded entirely before they actually leave the the govern
[29:24] environments. So again these violations are detected in real time and you can apply these controls automatically and every enforcement action is ultimately going to be recorded for audits as well that u Praep mentioned. So what this
[29:37] does is allow you to truly uh govern and manage how these agents are starting to uh engage with your data and focus on both the sensitive categories of data but also how your specific users and
[29:50] stakeholders are opting in um to their data and how it should be used by agents. So, same exact query here and what we're going to see is this is now going to return a subset of those initial records that were initially returned based on the subset of users
[30:04] that have actually consented to their data uh to be consumed and available for AI and um agent use. So, the vendor's AI interface now gets compliant results without any change to their workflows.
[30:16] And all this is enforced programmatically and automatically. Now, to to wrap it up too, what you'll see is the the proof here that it's um actually working. So the policy is live. It's scanning real data. It's catching and
[30:28] remediating violations as they happen. And you can see the trend down over six months. So what this is is this is governance that you can now measure and ultimately report back on.
[30:40] So that's one example of uh how we take risks and uh programmatically enforce those uh within data bricks. Now the next one going to focus on uh which you know was highlighted earlier
[30:53] today too. Data Bricks just shipped these LM guardrails as part of the Unity AI gateway. Now knowing which agents are missing which controls and and fixing that at scale that's a true governance
[31:05] problem and that's where one trust comes in. So every effective AI governance program starts with this level of visibility. You can't really govern what you can't see. So with one trust we automatically build and maintain this
[31:17] live AI inventory directly through multiple discovery channels. So in this example, one trust is connected directly to data bricks continuously discovering both AI models and agents as they're created um and deployed. Now once
[31:31] discovered the same policy engine can now continuously evaluate these assets against your organization's governance requirements and going back to the data bricks AI security um framework. So here
[31:43] what we're able to do is now set those controls for the enforcements in these specific guardrails that now exist within UNI AI gateway and can all be done um programmatically without necessarily requiring your AI engineer
[31:55] or technical stakeholder to go apply uh across uh the different agent ecosystems that you operate uh within. So rather than relying on these manual reviews and follow-up processes, we translate the governance risk into these actionable
[32:08] enforcement protocols that can then be automatically applied to the agents um at scale. So this violation is now automatically surfaced and then the user can trigger that remediation uh and is
[32:20] going to then apply the required guardrail on top of that agent. Now this change is going to be automatically applied and you can see it now within data bricks as well that these guardrails are are now intact creating this seamless connection between the
[32:33] governance policy and the technical controls. And now as soon as these guardrails are in place the policy engine is going to continuously re-evaluate these assets. So the violations now resolved the agent returns to a compliant state and then
[32:46] the inventory is updated in real time. And this is where one trust and data bricks help organizations move beyond simplifying uh and identifying this AI risk into really translating these governance requirements into
[32:59] programmatic enforcement um at scale. Awesome. So let me go ahead and bring this home. So AI is moving from advice to action. And these agents don't just answer, they act. And this action is now
[33:12] creating risk. So every agent that executes is a new surface area to govern and it requires runtime governance enforcement at the moment of action and
[33:24] agents are now making this urgent and we're seeing customers feel it. They're excited about this technology but they're also greatly exposed. They need the visibility into every model and agent. They need the speed to remediate.
[33:37] They need enforcement that they can prove. And then from AI inventory to runtime enforcements really need to invest in governance that helps to keep pace with the agents that each one of you all are building within your organization.
[33:49] So to conclude based on today's discussion um we'd be keen for each one of you to take the maturity survey to really get an understanding and place your organization on the AI governance maturity model. Uh we've created this
[34:02] assessment for you all to uh quickly take and you can share with your counterparts and leadership teams as well and really use it as a means to help benchmark your current state, help identify potential gaps and ultimately
[34:14] align on the priorities that are going to have the greatest impact on your responsible AI and agent uh adoption. So I want to thank each one of you uh for coming to the session today. I think if
[34:26] there's one thing to to leave you with these AI is moving at machine speed and governance absolutely has to to do the same. The governance and the organizations that combine that well are
[34:39] going to be the ones that are both innovating but again are intentional about this governance. And also before I wrap just want to extend a special thank you to PDep and the entire data bricks um team for the opportunity to be here
[34:51] in a great partnership and a great event these next couple days.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.