When was the 'Navigating EU and UK Compliance for AI Agent Identity Security' webinar held?
The webinar took place on 19 November 2026 at 10 AM ET / 3 PM GMT / 4 PM CET. (source)
The webinar took place on 19 November 2026 at 10 AM ET / 3 PM GMT / 4 PM CET. (source)
The webinar is designed for CISOs, security and compliance leaders, IAM/PAM professionals, security architects, platform engineers, and DevSecOps teams responsible for AI agent identities, privileged access, and auditability at organizations operating in the EU or UK, including US-headquartered companies. (source)
The webinar covered the EU AI Act, NIS2, DORA, and GDPR as they relate to AI agent identity security and enterprise compliance. (source)
Key requirements include assigning a clear identity to each AI agent, protecting credentials, limiting access to the current task, controlling permitted actions, enabling immediate revocation of authority, and preserving reliable records of sensitive activity. The Akeyless Identity Security Platform helps organizations meet these requirements by authenticating agents, establishing short-lived, policy-controlled connections, and providing runtime authority and detailed audit records. (source)
Organizations should identify which agents can access enterprise systems, replace embedded or persistent credentials with short-lived, policy-controlled access, apply least privilege, route connections through a controlled gateway, and for higher-risk actions, add runtime authorization, human approval, intervention controls, and detailed audit records. (source)
The EU AI Act's high-risk compliance deadlines have moved to December 2027 and August 2028. However, organizations operating in Europe must still comply with existing obligations under GDPR, NIS2, DORA, and other regulations. The delay does not eliminate the need to manage AI agent access, permissions, and auditability. (source)
Hi, everybody, and welcome to the webinar, to talk about the, regulation and what organizations are putting AI into production these days are facing in terms of, compliance requirements. My name is Yael Fainaro. I'm chief strategy officer at, Akeyless Security, and with me is, doctor Ilya Kolochenko, who is the vice chairman of the American Bar Association on the Information Security Committee. Hi, Ilya. Good morning. How are you? I'm very, very good. Thank you. Interesting topic we have, ahead of us. And before we, we jump into the, to the topic itself, just a brief reminder, everybody, we're gonna talk a bit about the EU and The UK regulatory reality, what enterprises should be doing today, and how Akeyless, can help you. We'll open it up for, q and a, towards the end, but feel free to type in questions as you, as you go. We're we're gonna look at them and and, hopefully tackle them as we as we go through. One housekeeping, note before we dive in, everything Eli and I Eli and I will share today is general information and based on current standards and regulation, not a formal legal advice. Please, consult your own legal and security advisers for any specific, architecture or decisions. Feel free again to drop, questions in the q and a. So, Ilya, let's start, with the regulatory picture. With the EU AI act deadlines being pushed a bit, what is the... What is currently governing AI agents in production today? So, basically speaking, if we talk about Europe or European Union and UK, I would say that... I would call this slide my favorite Big Mac or Cheeseburger Royale of regulations, whatever you prefer, because we have really a very complicated regulatory landscape. Let's talk very briefly about two directive. Many cybersecurity vendors claim that they help you to comply with NIS two directive, there is an important disclaimer here that only EU member states, Spain, for example, okay, or France can comply with NIS two directive because actually, in contrast, to, let's say, EU DORA or GDPR, that is actually a regulation that becomes law in all EU member states. Directive merely gives guidelines to the governments of EU member states to implement national laws to achieve certain goals and objectives. So basically speaking, you may have slightly different regulatory requirements based on the NIS two in France than in Spain or in Italy or in Lithuania, for example. That's an important thing to bear in mind. Another thing is that many people believe that NIS two is specifically to address cybersecurity, data protection and privacy in the private sector. Actually, have very small piece of legislation and IS to dedicate to the private sector, and it's mostly about very different things, nothing to to do with private sector. Okay? Very briefly about EU, Dora. I really enjoy this piece of legislation because it's very Wow. Enjoy. Wow. Yeah. Yeah. And I'll tell you why because GDPR, when we had it in 2018, it was like a panic everywhere because we had lot of interesting requirements like reasonable security and nobody really knew what it was supposed to mean. If you have a look at GDPR, currently, the only relevant article about cybersecurity, specifically, is article 32, that's it. And you have nothing about penetration testing, access control, security hardening, it's very high level legalese, okay? Practical suggestion, if you are willing to understand your legal duties, how to perform penetration testing or implement access control under GDPR. For example, you may wish to consult European Data Protection Board, advisories, and guidelines. They have amazing publications on many laws and regulations and only on penetration testing on GDPR requirement, they probably have a document that is 150 plus page long, so EDPB, and you'll find specifically what regulators expect from your side. Okay? EU Yeah. I would I would ask you, so none of these mention AI agents specifically. Right? I mean, none of them talk about AI as a as a special risk or as a dedicated risk, but talk about, like, operational resiliency, access control, auditability, and all of these things in more general terms. Whenever it's a machine or a human or AI agents, it's the same. Is that fair to say that all of these regulation apply to AI agents being deployed in an enterprise setting today? I'd say absolutely yes with some, I'd say, arguably narrow exceptions like, Cyber Resilience Act because it's about manufacturing goods that have some embedded electronics, Okay? For example. Alright? But, of course, you will sort of... Certainly have AI components, AI ingredients. So I would say, arguably, all of them apply to your AI infrastructure, whether it's in house, LLMs, fine tuned AI models coming from China, or OpenAI, it doesn't really matter. All of them do cover your duties relating to your AI infrastructure. Absolutely. So if if I'm, I'm an enterprise today and I have, somebody deployed, an AI agent, in my in my infrastructure, whether it was, permitted or not permitted, it accessed, let's say, data, and, and took out data that is GDPR, like, protect the sensitive data. Who's responsible for that? I would say there is no simple answer. And under GDPR, probably the safest answer will be data controller. Okay? But in a nutshell, I believe what is important to highlight that there is no excuse be it EU or UK, and even in The UK, we have English law, we have, different legislation in Scotland, so it's not that simple. Whatever your AI does, you cannot say AI did it. I have nothing to do with it. You will be liable, whether it will be, with GDPR or without. Just bear in mind that even if all these laws do not exist suddenly, okay, you will still be liable, accountable, and responsible for misconduct or damage caused by your AI and your AI agents. Under normal civil law. Right? I mean, not not even, dedicated cyber or privacy law, but just general, tort, contracts, like, normal civil law in every country. Yeah. You hit the nail right on the head that, basically speaking, you you know, in the eyes of law, whether it's your pet, like your cat or your dog, whether it's your car, your machine, your product, your software, your AI agent that causes some damage, you will be vicariously liable. Okay. So it seems like there's a lot that is already enforced. There's a lot of, chat... Like, there's a lot of different types of regulations, also local, like, we're talking here about Europe, but there's also in other countries around the world. So if I'm an enter... A multinational enterprise that operates in multiple places, I cannot check the law in every particular country that I operate to check that I'm meeting those specific requirements. Are there any principles, any guidelines, anything that I should make sure that I implement, and therefore, I'm most likely to be compliant with the current and future, regulatory requirements? I'd say that probably your strongest ally will be common sense. And as you currently share on this screen, when when you talk about global international compliance, it's virtually impossible to comply with every single piece of legislation, because in some countries it's sufficient to have one single client, and you must comply with all their regulatory duties and standards and everything else, okay? And another thing that even if you don't have any presence in any specific country, for example, you may have clients or partners or suppliers where you have a contract that still requires you to do, for example, visibility of your assets. And here, we have four pillars on the slide, visibility, verifiability, traceability, and explainability. So this share... Can you share a bit more about what each of those mean in practical terms? Sure. Sure. So visibility, this is actually, about, having a holistic overview and up to date, continually updated inventory of your IT assets, including your agents, AI agents, including their access permission, including lawful basis, why they may or may not access certain data, why they, are required to perform certain operations. Because if you use agentic AI, say, to process or update or securely delete personal data, You must have clear documentation, and you must understand what's going on. Okay? Why you are doing what you are doing, and you must also have logs and controls in place. Here, we can segue to the verifiability pillar. Basically speaking here, you need to make sure that you have security and privacy controls. And, for example, if you have an AI agent that is required by design to access your customer database, I mean, that's perfectly fine, but then make sure that that if you say that we need read only access, your AI agent will not start modifying your customer records, or it will not start modifying sensitive data or irrelevant data. So this is where you need to be able to verify that your agentic AI does exactly what expected, what documented, what plan. Okay? Because in my experience, when law enforcement or different administrative agencies, they start their investigations, to pay them administrative probe or criminal investigation, like we currently have in Arizona against the open AI for their security incidents with Agentic AI. Regulators usually ask you to provide evidence, meaning policies, procedures, and processes, and then they will verify that, check whether you actually comply with your own written policies. And be careful, I know that today there's a huge temptation to use chat GPT or any other tool to draft a state of the art policy saying we have everything, but in reality, have you have virtually nothing. This is the worst case scenario in my experience because it is better to have, like, foundational policy and have evidence of compliance, saying we are not superheroes, we cannot spend billions on cybersecurity, but we're doing our best, and we do at least the foundations, and we can prove it, then to have state of the art policy and and and and and and have countless gaps. Okay? Yeah. So something that can have teeth that can be, verified that it's it's actually implemented the policy, not just having on paper policy, which is really not giving you any any credibility with the, with the regulators if something happens. Exactly. Moreover, if you share, state of the art policies and you have nothing in place, you may even be, criminally prosecuted for fraud, misleading authorities, and many other bad things. So be careful. Make sure that your policies are actually being followed. Okay. So so this this covers the visibility and the ver... Very... Well, yeah. I love what legal is today. Yes. Yeah. Yes. So what's, what's traceability and explainability mean, with respect to the compliance? So these two concepts are actually closely interrelated, but, nonetheless, they are slightly different. Okay? Traceability, it's... For example, when something bad happens and, unfortunately, security incidents, have been happening, currently happen, and will happen. Okay? Yep. You need to have logs or any other evidence to investigate to understand what happened. For example, if your AI bot did something wrong or made an innocent mistake, you need to be able to investigate this because first question to answer to, for example, satisfy and discharge your duty to notify your customers in case of data breach, you need to know what happened, and I see plenty of cases when companies ring the bell and notify everybody that everything had been hacked, and later on, they make an update saying, actually, just 1% of our customers, will compromise, then only their first and last name, no passwords, no sensitive data. So be careful. Do not over disclose and do not under disclose because both scenarios can have very serious consequences. This is actually why you need to have traceability in order to understand what exactly happened, why happened, what's the compromised machines, for example, or who have been impacted among our customers. And explainability, this is purely outside about AI, and now it is addressed by numerous legislation on both sides of the Atlantic and Asia Pacific and Middle East and Africa and Latin America as well. When you need to explain why your AI decided to refuse a mortgage or why your AI decided to classify a transaction as fraud. So you need to be able to explain, and as a technical expert, I know that in some cases it's just impossible to fully explain from a to z, but at least you need to have some basic foundational explainability. Otherwise, you may be violating numerous laws and regulations. Well, I mean, I think the principles are clear. It, with, when we talk about autonomous AI agents, and I think the example of the Hugging Face, case that you you just shared shows how difficult it is to achieve those with autonomous AI. Right? Because they operate... They they kind of negotiate themselves out of guardrails. So you need to to make sure that that the guardrails and the controls you have are ones that cannot be bypassed, by by agents. And even if they do, kind of to understand their rational thinking, which is kind of a black box for us, as, as humans. Very clear. So, Ilya, what I wanted to ask you if you can share with us... I mean, we talked a bit about Hagging Face. You mentioned it. But Yeah. It seems like regulators are not waiting for any deadline. And and AI... As as AI progresses and as more cases are happening with daily breaches as you... As we can see, if you can share with us, a case, you don't have to mention the name or, something you, you have knowledge of where the regulators have found, companies liable for or close to finding companies liable for breaches of compliance regulation because they didn't have enough controls or good enough controls, in their access to information and and data. Sure. Sure. And probably just just very quickly to build on top of what you just said about hugging face. I mean, and without citing names, but when your agents, misbehave and break, you know, third parties and you start making huge marketing campaign on on it, then don't be surprised that, attorney general, sooner or later, will start scratching his or her head saying, well, let's investigate this this very interesting matter. Now, about Europe and The UK, I believe that we had some cases implicating artificial intelligence back in 2019, so before GPT, okay, when we had companies trying to blame them fairly trivial machine learning models, okay, saying listen, we do our incident response, we have our augmented SOC, okay, and we do respond to incidents as prioritized by AI. And actually we had a case in one of the EU countries where a company failed to disclose an incident in a timely manner, okay, and their main excuse was listen, we had purchased a state of the art AI sock management automation solution, and we've been just following AI guidelines, we believe that we reasonably believe and reasonably expect them to be reliable and we purchased one of the most expensive solutions from one of the most trustworthy vendors, so we've been following the guidelines. The court said, we don't care, you failed to detect because you didn't have any cross validation or cross verification that your AI does a good job, so you actually failed, you were negligent and the company had to pay a pretty large amount of money for a failure to timely disclose a data breach. So, we had cases back in 2019, today several examples with AI when AI goes off road, recent case study from my personal practice, obviously not citing any names, It was a European company, they decided to lay off a lot of people from their incident response team saying now AI will be orchestrating everything, AI powered SOC, AI powered future, AI powered miracles. So far so so good, and, AI can actually and does automate many, laborious and time consuming tasks. You you know, I've been always advocating for AI, but, like, you know, fair AI and, like, no miracle AI, don't mislead your customers, don't over promise about AI. And what happened, they started doing automated incident response with AI. In one case, hallucinated, to put it in a simple way, okay, hallucinated and started sending emails to a local law enforcement agency saying that their website has malicious content and in reality, there was no malicious content whatsoever, and started sending legal threats and started contacting all suppliers and all vendors of this law enforcement agency saying, please shut down infrastructure of this specific law enforcement agency because they are doing criminal activities. And one of the servers was shut down automatically by another AI system. When AI communicated to AI, they received an alert saying we have fraudulent activities and they automatically isolated potentially compromised environment. As you can imagine, later on they had very interesting and insightful conversation with the law enforcement agency in questions, and I believe you You should definitely know... Yeah. You should definitely choose your, you should definitely guide your AI agent not to mess with the law enforcement, agencies. Definitely, the case. Ilya, I wanted to ask you, you know, you talked about AI hallucination and things that are actually hard to, I think to to maybe, predict or control AI. What about, access to data? Like, access to a, to systems, more about the access control around keys, that encrypt the data or keys that, credentials and access control, the privileges. What what cases or what, what risk is there, from a enterprise perspective when it comes to, well, AI and not only AI, I would say? So I think this is one of the most foundational and important controls from the regulatory viewpoint because, you really need to ensure that you have holistic visibility of your users, their access permissions, their roles, their actions, their activities, everything. And I would say vast majority of data breaches and security incidents, however we classify them, happen because we have accounts that have overbrought permissions that do not have specific controls that limit access permissions, and by compromising one account in the cloud environment you can compromise virtually everything, or when, let's say, business team can access sensitive information and intellectual property of technical teams and vice versa. So I would say, identity and access management is probably one of the most important things to have in place, and it's not just like annual exercise. It's a continuous ongoing process to continually monitor, update, and revoke keys, permissions, and access controls. This is a very important thing. Okay. So, what I wanted to, to share or actually with you and and and get some of your feedback is we're the keyless. We are all about access control. And, actually, we have a a unique solution that also help companies, keep the, the sovereign aspect of keys because our keys are born distributed. So it's kind of you don't have it all in one place, and we don't have access to your keys that equip your data and your secrets. So it's, it also helps enterprises to meet the the... That aspect of regulation, which is really in in most... In many of the regulations. But in... With respect to the controls that you've highlighted, so the the way we approach things is we have a central place protected, actually, patent protected, platform that, gives central access, has central access policy, governance, and controls, really the highest possible, controls. So that, addresses the... And you can see what's, what identities, what secrets, what users, what systems they have access to, so kind of to satisfy the visibility and the ownership, aspect of the, regulation and, of course, the kind of full life cycle. So the fact that if somebody leaves, they don't have access and all of those things. We are providing customers, with effective controls, so the ability to remove anything that is standing. So hard coded credentials, permissions that are no longer... That should not be available to be abused. We're doing everything just in time and scoping to task so there couldn't be abuse of authority. We're checking every action before it happens, and that's very, very unique to AI agents because you can't predict their behavior. Right? So you you have to check everything they do, not just allow access and hope for the best because that we... We've seen how that ended up, in many, many, many, recent cases. And I think very, very important for our conversation, the the forensic, traceability and the audit. So because we are, essentially gating or kind of, all agent, workloads need to go through us, they need to identify, and then we give it permissions to ask... To act, which are very limited in scope. We can also log every action, not by the agent itself because we know that they manipulate the logs. Right? I mean, they change the logs, therefore, you you can't know what actually happened. And we can each attribute that to the original user. So you have ownership. So if somebody comes and says, like, in the beginning you said, oh, AI did it. A, you need to know if it was AI or not AI, and and we know that. But b, okay. So who who asked it to do it, and what was the the ask, and what was actually done? And the... To do that in real time so you can send an email to your customers with exactly what happened and what was impacted is something that we built our system from day one to do, and many of our customers use us, to meet compliance, requirements or the ones that are enforced, plus the ones that are, that are coming. So, that's kind of how we, we do that. Now before we wrap it up, I I wanted to, open the floor, to some questions. And there's one question that, that came up, which is to you, Ilya. So, practically, what's your recommendation for an enterprise, to do tomorrow? Like, what what should should we be doing, to make sure we're ready, make sure we are, we can defend ourselves if if something happened to build the, the ability to say to the regulators or at least not get to a to a case where... To a place where it gets to damages and and and court and and kind of be able to build the the the layers of of defense? So I think, once again, we should follow common sense because AI is here to stay, and AI will probably simplify and accelerate a lot of processes that currently consume a lot of time on our side, I'm talking about cybersecurity professionals, okay. AI will likewise accelerate cybercrime and cyberattacks. I don't think that we'll have any magic or miracle that AI will hug the world, no, but AI may streamline certain vector of attacks and it will probably amplify the damage, and basically speaking, exactly what you have on your current slide, talk about visibility and governance, I mean, with AI or without AI, these are the foundations and be 2026 or 2036, I believe that we are here to stay with visibility and governance with effective control that we need to continually monitor and test, Obviously, adjusting to take into consideration the evolution of AI agents, for example, because let's say in 2024, very few people work with Agenetic AI, now it's like one of the hottest topics, right? Maybe we'll see very different things in 2027 and 2028. And also, forensic audit, this is like something that is here to stay for decades, maybe for centuries, because we... With AI, with super intelligence or however we we call it, we'll still need foundational security controls, visibility, and auditability. So I'd say follow your experience, follow your common sense, and you'll probably comply with vast majority of laws and regulations. That's a that's a good good advice. That's what we have as humans still, common sense. Right? I mean, we should, we should use it until AI does that for us too. So it's a it's a... It was a pleasure, doctor Ilya Koluchenko, and I wanna thank everybody for joining us. If you have any additional questions, welcome to send it, our way. We can, obviously, address that, on a call or in in an email. So, Ilya, thank you very much for being with us today. It was a pleasure. And for everybody out there, good luck. Much for having me here to the day. It was a pleasure as well. Thank you. Thank you very much, and thanks, everybody. Thank you. Have a nice day.
The EU AI Act’s high-risk compliance deadlines have moved to December 2027 and August 2028. But the risks associated with AI agent access haven’t gone away.
For organizations operating in Europe, the delay doesn’t eliminate existing obligations under GDPR, NIS2, DORA, or other applicable regulations. As AI agents gain access to sensitive systems and data, organizations face critical questions: Who controls their access? What actions are they authorized to perform? And can those actions be traced and audited?
Watch cyber law expert Dr. Ilia Kolochenko and Akeyless Chief Strategy Officer Yael Fainaro explore what the revised EU AI Act timeline means for enterprises, how EU and UK regulatory expectations intersect with AI agent identity security, and the practical steps organizations can take now to strengthen access controls, runtime authorization, and auditability.
What do the revised deadlines mean for enterprises, and which obligations still need attention?
How do the EU AI Act, NIS2, DORA, and GDPR relate to agent access, permissions, and accountability?
Exploring least-privilege access, runtime authorization, and audit trails.
Practical steps to strengthen AI agent identity security, followed by audience Q&A.
CISOs, security and compliance leaders, IAM/PAM professionals, security architects, platform engineers, and DevSecOps teams responsible for AI agent identities, privileged access, and auditability at organizations operating in the EU or UK.

Cyber Law Expert

Chief Strategy Officer,
Akeyless
Trusted by Leading Enterprises, Investors, and Partners





