Loading summary
Transcript34 lines
- [00:05]
A
You're listening to the RSA Conference podcast, Where the world talks Security.
- [00:14]
B
Hello listeners. Welcome to this edition of our RSAC podcast series. Thank you for tuning in. I'm Tatiana Sanchez.
- [00:22]
C
And I'm Casey Zirkis and we are
- [00:24]
B
your RSAC podcast hosts. Casey, what are we going to discuss today?
- [00:29]
C
Well, Tatiana, this topic today is super interesting to me. AI powered impersonation and deepfakes, they're no longer theoretical. They are actually redefining trust in real time. We've seen this impact across the cybersecurity industry, from election security and even in the music industry. And that's why we're excited to be joined today by Clarissa Serta and Stephanie Fogel, who will discuss the evolution of deep fakes and share how industry and government can leverage existing legal frameworks, shared standards and regulatory intent to move from abstract concern to concrete action. Are you ready to dive in?
- [01:14]
D
Yes, I'm ready.
- [01:17]
B
Before we get started, we do want to remind our listeners that here at RSAC we host podcasts twice a month and we encourage you to subscribe, rate and review us on your preferred podcast app so that you can be notified when new tracks are posted. And now we would like to ask our guests to formally introduce themselves before we dive in. Clarissa, let's start with you.
- [01:37]
A
Hi, I'm Clarissa Certa. I'm the Chief Legal Officer at Pindrop. I work at the intersection of AI, deepfake detection and the law every day. And I've seen this challenge from every side in the White House at a global law firm and as a four time tech gc, helping companies bring novel technologies to market at scale.
- [01:59]
B
Thank you.
- [01:59]
A
And Stephanie.
- [02:01]
D
Yes. Hello everybody. I'm Stephanie Fogel. I am the Vice Chair of Global markets and sectors at DLA Piper and I am a partner in our regulatory and government affairs group. I work with global companies on regulatory compliance, crisis response and litigation preparedness, especially when technology, supply chain and enterprise risk collides. My focus is practical helping organization organizations navigate investigations and complex regulatory environments while keeping the business moving and on track.
- [02:37]
B
Well, thank you both for being here today to talk about deepfake. And it seems that deepfake technology seems to have industrialized identity fraud. So I would love for you both to explain to our listeners how these tools have evolved over the past couple of years to be so effective for fraudsters. Clarissa, can we start with you?
- [02:55]
A
Absolutely. Happy to start. Let me just dive right in. A couple of years ago when people talked about deepfakes, it was mostly about viral videos or political misinformation. Today, what we're seeing is it's about whether you can tell if the person on the other end of a video call is real. Think about it. If somebody can convincingly sound or look like you in under a minute, with today's technologies, a lot of our security assumptions start working. And if someone can deep fake you in real time, then every system built that's based on sounds right or looks right is exposed. That's the shift we're living through. I'm sure you're all aware of a roop. It's a widely reported case. An employee at a global engineering firm joined what appeared to be a routine executive video call. The faces matched his leadership team. The voices sounded right, the tone was urgent, but very believable. At the end of the call, he approved roughly $25 million and transfer was approved. And it was only afterward that they realized, here's the thing. Every single executive on that call was AI generated, was a deep fake, except for the one guy in the finance team. That wasn't fake content circulating online like we've seen in the last couple of years, that was synthetic presence manipulating a real employee in a real enterprise. And over the last years, deepfakes haven't just improved when they can do this, they've industrialized identity fraud. Voice and video synthesis is faster, it's interactive, it's scalable. You see it with deepfake executives authorizing transactions and synthetic candidates applying in live HR interviews. I'm going to mention just one other thing. Last summer alone, we saw over 15,000 unique AI generated voice bots in a single customer environment, collect information through the IVR and attempt to modify real account settings to drain funds. Those weren't abstract accounts that I'm talking about. Those were real people's savings, like fsa, LSA, HSA funds. The deep fake is what gets the attention, but what makes it feel real is that everything around it looks normal. The device, the location, the behavior. And when everything looks normal, people trust it. And that's what makes it dangerous. If identity can be manipulated, that easily stops being something you assume and it has to become something you design for.
- [05:54]
D
From a legal standpoint, given what Clarissa has been saying, evolution matters almost immediately. That's because the courts aren't evaluating whether the technology is innovative. They're simply applying the existing statutes like biometric laws, wiretap frameworks, privacy regulations. And they are asking very straightforward questions. Was there consent? When was their collection? Was their retention consistent with purpose? The courts and regulators are also very outcome focused. They will ask the basic questions were reasonable steps taken to protect against known threats. What harm occurred, how was the threat contained, and how was that harm mitigated? So as synthetic identity becomes scalable, enforcement expectations actually evolve with it.
- [06:56]
C
I'm not a lawyer and I'm trying to apply everything that you just said, Steph, to the example that Clarissa gave. And, you know, having taken reasonable steps in that deep fake situation where every executive in the meeting was deep faked, like, my gosh, that's a huge responsibility to be able to prove that you've taken the right, reasonable security steps to protect yourself from an incident like that.
- [07:28]
A
Right.
- [07:28]
C
And we're seeing this global push for online safety that, you know, the language being used is reasonable security standards. But, you know, it does beg the question, what are reasonable security standards that should be used to protect the entire enterprise, not just the network? Because obviously we've also got to protect our users from the legal and reputational fallout of a single deep fake.
- [07:59]
D
Yeah, Casey, you make a very, very good point. It is a challenge, I think, for all of us to maintain a certain level of security, given what has evolved in the world of deep fakes and fraud. So, as I mentioned earlier, regulators and legislators do love that word reasonable. It gives them great flexibility. But also, as you noted, it leaves organizations very uncertain into exactly what it is they're supposed to do. When you dig into that term reasonable, it's very contextual, what's going on around you, so the regulators will look at what risks were known at the time, what safeguards were available, and whether there was the right level of governance in place to address it. So, because deepfake impersonation is no longer speculative, and even at that time that Clarissa described, they were around, they were evolving and happening, they are considered to be measurable and must be part of the consideration of every company in assessing safety and in assessing reasonableness. So the real question becomes, did you take reasonable precautions given those circumstances? Did you address the risk? Did you evaluate the tools that could be used that you use? Did you document the decisions that you made and allocate the appropriate time and resources to very critical things to a business in thinking about what would need to be done to protect the company? So the real issue is defensibility. Defensibility, not perfection, is the standard here that will be applied.
- [09:49]
A
I'm going to pick up because you did a good job of covering the reasonable security, but the entire enterprise framing is important. This isn't just about a network intrusion, as we talked about it about. It's about hiring it's executive approvals, it's vendor payments, it's customer authentication. So security used to be about gates. Log in once and you're trusted. Now it's about signals. Are you still who you say you are? That's a very different architecture. Reasonable security today means your authenticity layer is, can withstand and handle these deep fakes. And something that often gets overlooked is that the friction doesn't come from the governance of what's happening and how people are trying to decide what this authenticity layer has to do. It comes from ambiguity. So at an enterprise level, if customers don't understand your safeguards, your deals are going to stall. If regulators don't understand your controls, your investigations are going to expand. So I think we need to think about it this way. Product innovation and innovation and technology that gets you live legitimacy means it works technically in the system and legally in the courtroom. And discipline is what keeps you operational and at scale. Does that make sense? It does.
- [11:29]
B
And you know, I had an interview with a subject matter expert and they were saying how there's been a rise in people going into Zoom meetings to get hired into a job, and they're using deepfake tools to make themselves look younger, older, different, you know, race. And it's usually hard for that HR person to detect because they may not have the tools. And when they talk to, you know, other employees who are interviewing that same person, they're not really talking about their appearance, their, you know, how they speak, but more so their skills. And we know that policymakers are already enforcing existing statutes against AI threats. But Clarissa, how can organizations use documented oversight and detection tools to define reasonableness and build a defensible position before a mandate arrives?
- [12:18]
A
Well, like you mentioned, I mean, you don't have to wait for new AI specific legislation. Although, by the way, there is a lot out there being drafted right now. Existing frameworks already apply. There are biometric statutes, privacy regimes, wiretap laws, regulatory oversight in different verticals, and shared standards like NIST and ISO. The structural tension is that most of these laws were written around discrete moments. Consent before collection, disclosure before recording, and AI driven identity systems operate continuously. So, as Steph said, you have to start thinking about how we define reasonableness now. Okay. And I think what that means is you start running explicit synthetic identity risk assessments, you're evaluating real detection capabilities, you're setting clear retention and model training defaults, you're building documented escalation and response purposes important for bringing a human in the loop when you need to. But most importantly, and I think the Most practical is you have to align your security team and your legal team's early in system design because everything is moving so quickly and that's what reasonable looks like in the AI era. It's not going to be a policy on paper, it's going to be alignment in practice. And when that alignment is missing, that's when enforcement usually shows up, right Steph?
- [14:00]
D
Exactly. Enforcement is very much focused on gaps and really does take advantage of and lay on top of the fact when there is not communication and alignment with security, governance and operations. So regulators won't hesitate to point to that as an example of what should have been done differently. And we see that time and time again as technology and companies and the way we do business evolves. So for this reason, preparation is super critical. Regulators will ask for documentation first. They're not interested in your marketing materials. They want to see the documentation. Did you evaluate the risk, did you consider safeguards? And again, most importantly, was there cross cross functional alignment? And the record really shapes how reasonable is going to be interpreted. Calculated preparation and analysis of these threats before an incident occurs is what is going to set the tone for moving forward.
- [15:07]
C
Stephanie, you had mentioned security used to be about the network and you log in once you get through one gate and you're good. And we've clearly moved from the one time gate to continuous AI signals. And I appreciate you both bringing up alignment because my next question is having to do with that. So how do we align these always on systems with event based consent laws to ensure our identity verification remains legally defensible?
- [15:46]
D
Another great question. I'll answer that one. The event based consensus laws weren't written for these always on systems, which is not surprising. Yet they do continue to apply. So that means that institutional and intentional design choices need to be made. There should be clear notice. And enterprise configuration controls, defined retention limits, always human oversight and technical robustness. Compliance by design is really critical to this evolution and can help to systematize safeguards and reduce manual failure. And again, the courts analyze and are looking at whether statutory requirements were followed, not whether the product was innovative. So design choice is also central to legal defensibility.
- [16:40]
A
And Steph, just jumping in. I mean, I think the consent question. Look, consent is going to remain foundational, right? It's what gives the customer trust. But in systems that continuously evaluate identity signals, consent alone isn't enough. Identity is being evaluated continuously. So identity assurance and legitimacy has to be continuous too. It has to be visible, it has to be documented and it has to be enforceable in time let's take a look at what we learned from the or what we're learning because it's not finished. The Otter litigation. It was a tool designed to make meetings easier. Quietly auto joined calls, captured people who never signed up and allegedly used conversations to train models. Overnight, a product decision which was meant to help and it was about a frictionless experience, became a story. And a litigation about secret recording and wiretap risk. That's the moment we have to design against. For a CISO who's building this stuff, it means building purpose limits, proportional safeguards and audit trails into the architecture from day one, so security, privacy and user confidence can scale together instead of colliding later. This isn't a lesson to slow innovation, it's just to remind us that we should shift capability and accountability together. If identity can be manufactured at scale, identity assurance can't be an afterthought.
- [18:33]
C
Clarissa, I want to just dig into that a little bit because it's fascinating, right? And maybe it's that we have the gift of hindsight to be able to look back and say, like, how did they not see that coming? But how do organizations, developers solve for these sorts of. Oh, I never considered that secret wiretapping, right? Who should be a part of all of these conversations? Because the question is about alignment, right? And who within the organization, from the developer to the CISO to the customer, needs to be aligned to ensure that they're mitigating that risk of legal action in the consequence?
- [19:21]
A
I really think you have a couple of things working, right? One is, can the system work to catch these things that are coming in? Is the system going to be legally defensible because you're doing the right things with whatever the legal frameworks are that you're subject to. And third, is what you're doing going to be something that's transparent enough that it gives trust to the people who need to use it, and it's interpreted in the vein of being for their safety rather than surveillance, for example. So to answer your question directly, though, to move from abstract concern to concrete action, my recommendation is you start treating authenticity as infrastructure. You start with a trust inventory. Where are we assuming identity authenticity? You look at places like hiring interviews, executive approvals, your video, teleconferencing, your vendor payments, you know, things like that. You stress test those assumptions. I'm a big fan of running a tabletop exercise and see if you can run one where someone attempts to deep fake an executive or pretend to be an HR candidate and see how far they get. You find the Gaps, you fix them and you test again. Here's the thing that a lot of people often forget. If you don't rehearse this, sometimes this feels really funny. But if you don't rehearse this, you're going to be figuring it out for the first time in the middle of a crisis. And these. Right, you don't want that to happen. And we're at a point right now, just a little reminder, we've moved from fake content to fake people. And when that happens, the victims aren't just headlines, they're individuals. This stuff is really starting to matter. So organizations that treat authenticity as infrastructure and not just compliance are less likely to be explaining themselves after the fact. In the AI era, the ability to distinguish real from synthetic in real time becomes a core control. And if that control isn't transparent and defensible, it won't earn trust. And if it doesn't earn trust, it won't scale. Bottom line, adaptive security requires adaptive legitimacy.
- [21:54]
B
And Clarissa and Stephanie, you guys both mentioned, you know, compliance by design. And it also reminds me of being resilient by design and anticipating those cyber attacks, deep fakes, you know, the looking at possible outcomes of if, you know, a hacker can do this or that, how can we protect ourselves and implement security tools to defend against them? And it's often said that governance failure is often identity failure. So, Stephanie, how can organizations move away from just checking off a simple checklist and then leaving it at that? And how can they involve their crisis playbooks to specifically anticipate synthetic identity failures?
- [22:34]
D
Well, I think that in institutionalizing and integrating ongoing and emerging threats, companies simply need to update the playbooks and the response tactics that they have in place. But what I think is most important to think about as move into this new world is that many of the things that we have always done were very binary or very linear. And now we need to think in terms of a much broader type of multiverse where multiple things are happening at the same time. And frankly, we need to start to integrate AI, such as the things that companies like Pindrop have done, into the way that we respond to these types of attacks. So it's imperative that companies are not relying only on a human ability to follow a checklist, but to have actually integrated AI in a successful way to address bad AI. So those checklists and those playbooks and those response tactics and those teams become multidimensional and are much more effective. And there's a certain level of communication, transparency, and immediacy that can only be accomplished in that way. So it's a rethinking of the way we address problems in just about everything that we do.
- [24:07]
A
So I second everything Steph said. But I would look at it this way and I would say identity failures are not things that sort of happen in isolation. Usually when they happen, what has happened is that your assumptions that you built your system on weren't stress tested. So that's part of the reason I advocate for the stress testing above. Right. I mean, at the end of the day for me, I think some of these failures are governance gaps that were revealed under pressure. And this is a whole new world in terms of what is happening with this continuous layer. You think about it this way, if security has to become, and there's Jim Ruth out there who is a very well known ciso, has said that the way the world is going and where technology is right now, security and the way you look at it has to become almost like your body's immune system. Right. It has to be able to respond. It doesn't have time to do stuff. And so I agree. And if you have that kind of environment, there's no time to do the old school, let's go do these checks after the fact and let's involve. So I think all the things we've talked about culminate in making sure to make your systems be where they need to be. You have to be cognizant of the fact that they're going to be moving much faster. They're going to have weak points you didn't think you had before because of the advancement in technology. So we need to make sure that the defensibility in the system, both in the system and legally defensible, is taking into account that speed at which it has to interact, all the different places where it has to interact. And the only way you're going to get there is by making everybody smarter. And your security partners have to understand where they can push when they need to, where they don't have to. So really having that alignment up front in terms of guardrails and what you do and what you don't do and how you handle it is going to be super important. Because if you wait to have to be able to check with somebody after the fact, this stuff is just going to happen more and more. And so I just think there's a need for all of us to work together and be way more aligned to get ahead of this.
- [26:38]
D
I agree.
- [26:39]
C
And I think the alignment needs to come also from a development standpoint too.
- [26:45]
A
Right.
- [26:45]
C
Because the examples that we talked about like the tabletop exercises and the threat modeling shouldn't only be in response to an external threat but potentially creating an internal threat with what we're developing. So being able to do those tabletop exercises and have that alignment across the whole cross functional alignment within the organization before releasing products is so important too. I am pleased to see legislators in Utah and across the states pushing back for AI regulations and hope that we continue to move that conversation forward. Stephanie and Clarissa really appreciate you being here today. Great perspective on this and one that I'm sure we'll continue to revisit as things evolve. So thank you so much listeners. Thank you for tuning in. Yes, please come back and join us again listeners. Thank you for tuning in. Please keep the conversation going in our RSAC membership platform by visiting onersac.commembership and be sure to check onersac.com for new content posted year round. Finally, don't forget to Register for RSAC 2026 conference by visiting RSACconference.com us until next time.