The vulnpocalypse might not be so bad after all Transcript, AI Summary & Key Points

00:01 Depending on who you ask, the AIdriven Vong apocalypse is either the end of cyber security as we know it or it's a load of hot air. Uh panelist, what side do you land on? >> Probably I'm landing on somewhere in between. Um I don't think it's either, but I'd say closer to the hot air. >> I agree. Somewhere in the middle. Uh maybe I'm being overly optimistic, but I think that in the long run, uh it's a good thing for the industry >> as we know it.

00:27 ES. the end of waiting for things to break before patching it, but it'll be good for cyber security overall. Okay, I like that. That's a good one. Hello and welcome to Security Intelligence, IBM's weekly cyber security podcast where our experts turn the biggest industry news stories into practical takeaways that you can use. I'm your host, Patrick Lucas Austin, and joining me this week, Gakamo Cassone, senior exhorse incident response consultant.

00:56 Brad Lair, security consultant and Norman Dorsch, product manager at X-Force. Uh, plus later in the show, uh, Schweda Jane, head of prominatory IBM consulting, drops by to talk about how we have to redefine the cyber resilience in the age of quantum computing and AI powered threats. Before all that, we're talking about AI agents turning unsuspecting wikis into their own private message boards and why cyber criminals are sticking with the tried andrue tactics in the face of all this AI upheaval.

01:27 Uh but first uh we are here to rethink the V apocalypse. AI has opened the vulnerability floodgates, contributing to back-to back record-breaking patch Tuesdays over the last year. As I'm sure you've all seen, um, cyber security professionals are really wondering, how are we going to fix all of these vulnerabilities and a new report um, suggests that we might not have to fix them all.

01:58 We may just have to double check them and uh, and validate them. Um, we've got a we've got a mythos readiness report um that ECHKO has analyzed over 40,000 CVES and found that while most CVE closures have skyrocketed, a lot of them are a lot less serious than they um, you know, than are initially believed. So I think the question for you guys is what do we think about um focusing on validation rather than you know going out and and rushing to solve these so-called critical vulnerabilities.

02:32 Uh Brad, let's go to you. >> The first thing I would say is just because something's been deemed a vulnerability, the first question should be does this apply to me and what does it mean for my business? And then just because an AI an AI large language model in this case mythos says it's a severity of 10. Well, is it actually a 10? So, it's important to trust, but you also need to verify.

02:55 I don't want to pick all the meat off the bones for the other two gentlemen here, but I'll begin with that. Agree with uh Brad there. And I mean, Brad, you were you were alluding to the business context, right? Like, what does it mean for me? And I think that's the most important part. And I think the report cited one in eight of the critical findings were not really deemed critical by the actual humans reviewing it which is precisely the point right so we got to look at okay what does it mean for the organization um

03:25 and then often yeah that's where we stand on the or when we get back to the question is it all hot air or is it the end of uh everything it's probably somewhere in between and and we got to confront the fact that we have a um vulnerability crisis, yes, but we also have a remediation crisis. And I think that's what's really been exposed. There's a remediation crisis.

03:52 Vulnerabilities are out there for multiple months being unpatched. Uh but the patches are available. So that's that's the bigger issue, >> right? Um the report points out that a lot of these vulnerabilities are are self-inflicted. They already have fixes available. they are they can be solved with basic cyber security hygiene. Uh Jakimo, what do you think about the the sort of rush to use AI to find these vulnerabilities at at machine speed but sort of you know fix them at at human speed often kind of you know slower

04:24 since we're since we're waiting through all of these you know quote unquote critical CVEes. >> Yeah, I think I think potentially uh we go back to what Brad said. um the role of vulnerability research in in the field of security uh has probably changing somewhat. Uh I think finding the vulnerabilities is somewhat less valuable cuz it's it's it's a bit easier.

04:45 uh and so I think that the role of vulnerability research should be moving a little bit more towards uh remediation right so well here's a vulnerability whether it's it's critical or not that obviously depends on the b business context uh and where where it is right um but then also try and come up with suggested fixes um which isn't just a patch right is how do you actually fix it in a in a real business context um and And I think that's that's part of what we're going to see uh moving forward um in in in terms of

05:21 changes in uh vulnerability research. >> Is there any way to sort of validate at the same rate we're getting these these vulnerabilities are sort of coming in. How do we apply that same that same speed to fixing the problem? >> It's not necessarily about my my perspective on it isn't isn't necessarily about validating. It would be how do vulnerability teams work with the architects to make our attack surface small enough that we actually know where our vulnerabilities are and we're always monitoring it because it's the

05:52 unknown unknowns that people aren't going to patch and that's what's going to hurt them. Garbage in, garbage out. Yes, exactly that. >> That's that's a that's a very good uh way to uh hop off of this uh first segment and on to the next one. Um but for our viewers on YouTube, you can uh leave a comment and uh let us know what you think about the wrong apocalypse.

06:12 how your organization is handling it and how you're finding it challenging or maybe not so challenging. Rogue AI agents using websites as private message boards. Um, researchers have identified more than 10 instances of open AI related AI agents secretly turning public websites into message boards to share information and coordinate activity with one another.

06:38 It's not something we haven't heard before. Um, recently OpenAI said this happened a few weeks ago during a during a test where they it uh it's uh agents escaped their sandbox. Um, and we're we're back at it again. Um, OpenAI hasn't said what the agents were doing on these message boards, but researchers theorized that they were looking to circumvent restrictions to cheat on yet another assessment task.

07:01 These guys would have been great in high school. Uh, questions for the panel. What do we make of this? AI agents taking over uh very obscure wikis and websites. One uh one investigator found traces of activity on a high school teachers um wiki setup in 2008, which is in incredibly, you know, obscure. Um what do you guys make of all this agents taking over out of the way websites using it to concoct their schemes and cheat on tests?

07:34 Is this something we should be worried about? Uh Jakamo, how about you? >> Yeah, overall I I think so. Um the main concern in in this case uh which I think is really interesting uh the story is that um it's pretty understood or sort of accepted right that when you talk about security um in agentic applications that the security boundary should live somewhere outside of of the agent and the LM because you can't enforce the the security at the LLM level.

08:02 Uh but I think here we see an example of where this A open I tried to do this. Uh but you know when the when the rubber meets the road this actually is not so easy to do in practice, right? Um you hear the examples about yeah well you want to give your agent low privilege API key scoped only whatever it is they they they should be able to do but in this case OpenAI wanted to give internet access uh but wanted to give only read only internet access.

08:30 And how do you achieve that? Um, obviously they try to drop state changing requests, right? Like post, put, delete. Um, but this is the internet. Everybody does whatever they want. Uh, and so then that's why the agents managed to figure out that these obscure websites could be uh tampered with. Um, and so I think this is a very good example. Uh, it's scary because it's a is a good example of how we have a theory about how we should secure these agentic applications.

08:58 Uh but in practice it's clear that we're not quite ready um to actually put these guard rails um and make them really bulletproof. >> Sure. Openai says it is working on a a framework for reporting uh misalignment. Um Brad, how effective do you think that's going to be? That looks like a face that thinks it's going to be very effective. >> It could be.

09:21 However, whenever you have an autonomous agent and has agency, you're going to have a really hard time getting trans transparency and visibility into everything. And in this case, even if you say real read only, maybe you didn't say you're not allowed to use uh chod and change your permissions and they mistook that. Maybe it wasn't explicit enough. Maybe they just ignored it.

09:41 So, I think it's a great idea. I'm not sure how you enforce it, though. Just because something's a rule, that's one thing. Can you make can you make them do it? It's the the I guess the clever, you know, nature of this of this hidden activity makes me wonder how many more agents are out there uh that we haven't yet identified, you know, plugging away at at, you know, exercises um trying to find workarounds around them.

10:09 Norman, should I be should I be wondering if there's an agent, you know, hiding in my in my cup of coffee doing some math? Not yet, but eventually. Yes. And I think the like how many agents are out there uh doing things they're not supposed to do. That's the million-dollar question, right? Or the I guess the trillion dollar question. the um I mean what happened in that exact case that the researchers were looking into was they had readonly permission but found a way to edit comments and wiki pages which as uh Yakimo

10:42 was saying right that didn't um uh use uh post um protocols in in uh over the browser so we put those guard rails around the agents when giving them those handboxing test, but we didn't test for their actual execution and we didn't weigh in the creativity that those agents are now having. And so we didn't monitor any of the uh improvisation that they are doing when the task is set.

11:13 The task the goal is their ultimate goal that they want to do. And uh they're trying to do whatever it takes to get there. And I think that's what's very worrisome. Um the creators of those models now see that too. I mean uh Daario Amade just published an essay saying hey we should all uh relax a little bit and and um slow down a little bit and uh Altman agreed, Musk agreed.

11:44 So they're they are recognizing this and they are pushing themselves for a for a slower pace um and more regulation which is otherwise unheard of from private companies. I mean the the desire to integrate more AI agents into cyber security workflows makes me wonder can we trust these agents to operate uh more independently in these workflows if they are going to look at the guard rails look at these these rules that we've set and say I'm going to maybe try to work around this because it's kind of getting in the way of

12:18 me reaching the my end goal >> at the AI agent level it becomes an identity provision problem. How do you explicitly define AI agent one can do these set of rules and cannot move out of scope? It's not enough to say don't. You have to make it such that it can't be done. >> How do you monitor it once it's once it's deployed? What's happening at runtime?

12:39 It's one thing to set the guardrails in the theory and set it during the the setup of the of the capture the flag or the real-time engagement, but then who monitors the actual execution of it. And with hop tooling cuz this is another interesting point. I think we normally uh in security monitoring uh we're used to structured data logs telemetry um but then all of a sudden we have all this natural language that is being produced by this agent and that is this control flow essentially right it contains intentions it

13:11 contains you know to-do list things like that. um how do you analyze it without using another LM which is susceptible obviously to the same issues it's like a sort of oraoros of AI our final story uh a tried andrue cyber security classic vishing uh shiny hunters have turned to it to uh exort the the one of the more vulnerable industries the healthcare industry We talk about how AI is changing cyber security, but hey, old school tactics still work apparently.

13:50 Um, the threat intelligence group Health Isac introduced a bulletin warning hospital systems that shiny hunters are using voice fishing to harvest credentials uh that hold sensitive health data in exchange for ransom. Um, you've you've heard this before. The operation starts with a little recon. the uh bad actors figure out who they want to impersonate, usually IT help desk.

14:16 Um they then they call employees directly on their devices using um using uh voice fishing. They uh send follow-up voicemails and emails asking for particular bits of information and that leads to of course a cyber security compromise. Um the likes we are very familiar with. Uh I'm going to ask the panel. Fishing is very wellnown. It's very common. It's a very annoying social engineering tactic.

14:44 Um, why is it why does it still work? Why haven't we figured out uh that we need to call the person back and uh just, you know, double check that they're really asking for their social security number? Why don't they have their own social security number? Uh, Jakamo, >> uh, with all this talk about AI and AI enabled attack, people may be forgetting that, um, there's no AI agent out there doing anything crazy, uh, that we haven't seen before, right?

15:14 Like fishing. Uh, and so the companies and organizations in general need to keep sticking to the fundamentals like uh, uh, security training, but the usual, right? Uh, which I guess ties in back to our um our first point. We may be worried about like a apocalypse of vulnerability of critical severity that are going to be raining down upon us. Uh, but really uh the real threat I think is u the the usual attacks um the same avenues that have always been used just a little bit quicker.

15:50 And so organizations really need to keep working on the fundamentals uh to stop AI and nonAI attacks. >> Norman, do you think that uh do you think it's uh that much of a surprise to you that this is uh still working despite the surge in AI powered uh attacks that we're seeing? >> No, it's not a surprise. I mean, at the end of the day, attackers are pragmatic people, right?

16:13 They're they're using what works. And if the old school uh fishing attack works, why not do that? And uh you asked earlier why is it so powerful, right? I mean it's just anything social engineering related always has the human aspect on it and that psychology aspect that whole sense of urgency. You want to be nice to the person on the phone. You don't want to be messing up with your help desk or your manager gets notified.

16:37 So you are going to uh in this case I think they read out the the MFA tokens that they were sent like the onetime um passwords. So that human aspect is still one very breakable um point in the in the uh in the attackers's uh methodologies and that whole attack there just proved I guess I mean the whole your if the identity becomes the perimeter and and network doesn't anything you've been doing on network for ransomware preparation all the like segmentation all the downtime playbooks we always do.

17:17 None of that matters. It's all irrelevant when you bypass at the identity layer and if that be becomes the perimeter and then you got the lateral movement from there. So, um no surprise. Sorry for the long answer but uh doesn't surprise and it's it's a good way to that still works unfortunately. It would have been preventable though like um there are measures of course so like moving to um different MFA methods uh pass keys would not work through that particular uh attack path.

17:49 So there are ways >> yeah my uh my my followup is sort of how how much progress have we made um from a cyber security perspective combating things like voice fishing and other and other sort of fishing attacks. Brad, do you think we are getting better at it or is there just more of it to deal with? >> I would say there's been a bit of a stagnation in implementing the basics like what Norman was saying, we already know what the solution to MFA is.

18:19 You could make it so that change it from something you know to something you have and it's a physical security key or what have you. And you can also do other things like once once they were getting in there, they were pivoting on the identity dashboards. So the lateral movement was what they were really aiming for. And the social engineering just happened to be the most expedient route.

18:40 So we could secure it such that if they do get in, they can't see anything. And every time they have to enter a system, it doesn't give them anything. And you secure it right at the identity and make sure the permissions are audited per action. We know what the answers are. It's are they going to is it will it be implemented? >> I think that's a that's a great place to end it.

19:00 Um, thank you so much Norman, Jakamo, Brad for, uh, coming on Security Intelligence. Uh, it's been a joy to have you guys. >> Thank you for having us. >> Thank you for having us. >> Pleasure is mine. >> Up next, Schua Jane joins us to talk about a new article she co-authored with Steven Karajio. The next cyber crisis is already taking shape. The piece explores how advances in quantum computing and the AI powered threats are combining to reshape the meaning of cyber resilience today, especially in the financial sector.

19:32 Sha, thanks for being here. >> Thank you for having me, Patrick. It's a very timely topic and I'm excited to be speaking about this uh article that we co-authored. >> Your piece highlights the confluence of two factors, thread actors gaining access to sophisticated AI tools and banks transitioning to postquantum cryptography. Can you tell us a little bit about the dynamic this creates in the cyber threat landscape?

19:56 >> Thank you, Patrick. It's a timely question and you know before I sort of launch into the answer for this question I would broadly say that the key point emphasized in this article is really that the risk is not just stronger attackers or new cryptography in isolation. It is the convergence of both attackers are getting faster and more capable um at the exact moment banks are undertaking a fragile multi-year rebuild of the cryptographic foundation that underpins trust payments identity and operations.

20:26 So it's similar to basically saying that the locks are being changed just as the lock picks are getting better. So in the context of that objective when I think about your question at the heart of it the real issue that we are discussing is asymmetry. Banks are in the middle of generally speaking complex multi-year cryptographic transition and everybody is at varying points in that journey while attackers are getting AI gaining AI tools that let them to find weaknesses and scale attacks much faster.

20:56 So the risk is not just stronger attackers or a harder migration and isolation. It's really the convergence of the two happening at the same time that creates a period where financial sector is more exposed to more predictable and dire threats. Um the two trends reinforce each other in a relatively challenging way. AI is lowering the barrier to advanced offensive cyber capabilities and at the same time PQC migration is forcing banks to touch deeply embedded cryptographic dependencies across thousands of systems, APIs

21:30 and processes which basically means that the attack surface can expand precisely when adversaries are becoming more capable. So this is not a theoretical future state issue we're talking about. We're really talking about something that is happening in the here and now. Historically, the most sophisticated exploit development was concentrated amongst a relatively smaller number of actors.

21:52 That has changed with Frontier AI. Frontier AI starts to democratize that capability. So the concern is not just more attacks but faster, more scalable and less predictable attacks and that is where PQC migration is essential but it is operationally very disruptive. Cryptography is not a single product you you know can swap out. It is woven into the customer authentication layer into the payments into data protection and digital channels.

22:18 And from a risk perspective, this presents a timing problem where critical cryptographic dependencies sit and where services are most vulnerable, where third party risk is being managed and really where the compliance question becomes, can an institution evidence a credible plan, an accountable governance for mitigating these type of risks. And here I want to end with the key point that what changes here is not simply the threat level.

22:46 It's the mismatch in the speed. Attackers are speeding up while banks are in the middle of infrastructure transitions and other enterprise level plans. And that's really where the strategic risk sits. >> Thank you. uh your piece argues uh that we need to rethink resilience that we should approach it um like you're saying as a sort of managed degradation rather than a recovery exercise since all of these different um elements are are at different stages of of development or degradation.

23:14 Can you expand on that? Um what does that mean and why is the why is that the response to this particular moment when these two major um uh changes are coming together? Great question and and I do want to acknowledge that resiliency has been at the heart of really what I have spent a lot of my focus on with my practice because really in my opinion we need to move beyond the idea that resilience means restoring everything quickly back to normaly in a sustained cyber event that may not be realistic.

23:47 Managed degradation means deciding in advance which services must survive and which can be curtailed. So you're making very proactive risk um you know tier decisions and how to preserve core functions such as payment settlements and liquidity under stress. That is a much more useful definition of resiliency currently rather than saying that everything will return back to normaly and you know normal um return times.

24:11 Traditional recovery thinking is too too static um the way we traditionally have thought about it. It is built around recovery time objectives predefined failovers and a return to normaly. But that model assumes that the institution is recovering from a contained event. It is less useful when the institution is going under attack with multiple dependenc dependencies stressed at once.

24:32 Managed degra degradation means deliberate choices under pressure. And that is really where the bank governance accountability and their systematic way of prioritizing risks um and processes really comes into play. The questions to ask are which services do we need to preserve? What is critical for the business? What is critical for the future strategic planning?

24:54 Which services can be throttled and restricted or temporarily suspended or be managed with compensating controls? And while doing so, what are the core operations such as clearing, settlement, liquidity, and customer protection which really require that fundamental focus. This basically makes it more of a governance issue before it becomes a technical issue.

25:15 you really need to get ahead of these threats and determine which of these processes really require that type of um uh uh you know recovery time and how do you bake that into your plans. So therefore boards and risk officers you know need the right level of governance decision rights and escalation thresholds for management of these type of situations and there should be clarity on what core services really really entail.

25:42 Um and in that from that perspective the objective is shifts from uptime to trust prevention. It really is about balance sheet and trust preservation when failure is already occurring. It really takes out resiliency from just the technical realm to overall enterprise resiliency and looking at what it would take to manage um not only regulatory and market um expectations but really the fundamental of the business on what will survive versus what will need to wait.

26:12 And in this case, you know, I would recommend thinking about this in basically the following way. Resilience today is less about a clean recovery and more about control continuity. In other words, not pretending that nothing will fail, but deciding in advance how to fail safely and how to manage that in a very explainable way. >> I want to talk about uh postquantum cryptography next.

26:37 Um I know it's going to go a long way towards securing financial institutions against threats, but I it doesn't stop there. Um, so what else should banks be thinking about when it comes to shoring up their cyber defenses for the future against these, you know, new threats like uh these frontier models that are coming out? >> Sure. I mean, look, I think postquantum cryptography is definitely front and center and a key area of focus, especially with the advancement that we're seeing, and it is necessary, but in my

27:05 opinion, it is only one part of the answer. Banks also need full visibility into into the cryptographic dependencies, stronger vulnerability management, better thirdparty oversight and resilience plans that are tied to critical business services just like we talked about in the prior question. So basically, you're creating your own prioritization plan based on what is important for your risk profile in a very explainable way that prioritizes the right processes and the right defenses that you need.

27:34 The future defense model is not just about better encryption. It is about better governance, better prioritization and better execution um across the enterprise. You can start certainly with the cryptographic visibility. Um and you cannot protect what has not been invenied. You need to know what exists. So it really begins with a clear map of where cryptography is being used, which algorithms and certificates are being deployed, which applications and vendors are relying on them and and ultimately which business

28:03 services would be affected by these type of failures or delays. In most institutions in my you know experience the real risk has not materialized in cryptography itself but in the hidden dependencies and the dependencies underlying within the risks that go generally not I would say not mitigated but overlooked because they are put under the guise of cryptography and therefore not interconnected to that level of planning that we were talking about before.

28:33 So from that perspective, my recommendation in this case is always to say that prioritize by criticality, not by convenience. So not every system will move first. It should be again going back to the fundamentals of what are your core services, what are your core systems, how are you managing your identity and access management, where is your customer data living and how is your market infrastructure connectivity set up so that you can then evaluate you know where those touch points may be.

29:02 critical piece of that is also thirdparty integration. Um spend a lot of time looking at thirdparty risk. So having uh more you know uh risk based migration is more defensible in my view compared to a one-sizefits-all. Um equally important is you know strengthening software and vulnerability management and that remains a big area of focus both in the um article as we describe it but also in current discussion topics.

29:27 This entails everything from having the right secure development uh and patch prioritization and code reviews, having the right open source governance more so than now, especially given all the news um that we read today on AI governance um and attack surface management and you know especially important because the article itself highlights increasing concerns around AI assisted vulnerability discovery and exploit chaining.

29:52 So, you know, bringing this all back together and thinking about building cyber resilience into business and treasury thinking and improving decision making. Um, really it is about having a foundational control PQC as a foundational control but not you know just a strategy. The real strategy is knowing your dependencies, prioritizing appropriately and making resiliency executable across the enterprise, not just uh within the security team.

30:24 >> Are there any last uh words or or final takeaways for for listeners? Sweda, if I'm a cyber defender in banking, what should I start doing? What's what's my next move? >> Well, you've asked for a little bit of a playbook. So, you know, in my view, it's it's what I'm going to outline is going to be very simple. Um and you know the greatest words I've heard are anything that is simple is generally the hardest to do.

30:46 That I absolutely think is the case here. Um where you know it is very simple to say identify the business services that truly matter. Map the dependencies underlying those processes and those services and prioritize action there first. It really is that simplistic. However, getting a little bit more nuanced. We have talked about how the speed at which these threats are materializing is a key variable in this external variable in this and how governance comes together particularly in complex enterprises with many

31:21 competing priorities and many competing frameworks this can be a particular challenge. So in those cases, you know, I I would recommend that really starting with that inventory of services that matter the most, having that full enterprisewide view and connecting into your BCM disaster recovery and resiliency programs in a very meaningful way um across the three lines of defense really and then mapping the dependencies in the same way across applications, certificates, encryption libraries, APIs, your network

31:49 components, your vendors and really all of your processes. And then fundamentally asking three very hard questions which are where is there exposure to cryptography? Where are we cryptographically exposed? Where is the enterprise operationally fragile? Where do we have heightened risk? And then finally if there were you know a sustained attack what would be deliberately u um degrade to preserve the core?

32:19 what what is our plan to meaningfully fail so that we can control and manage that the best of enterprise ability and running those scenarios now not waiting for the um attack to you know materialize not having generic tabletop exercises but severe realistic scenarios which involve accelerated exploiting activity vendor weaknesses crypto dependencies and prolonged disruptions of digital channels and then having the right escalation paths and evidence to show that this type of prudent risk management is occurring.

32:53 I think overall I would close by saying that in my opinion the biggest mistake now is mistaking the absence of a fixed deadline for the absence of urgency. Particularly as we await regulatory guidance as we await which way the administration is going to you know basically lay out the framework as again we heard in the news to this just this morning in banking time lost in preparation is usually uh recovered later at a much higher price.

33:22 So that preparation truly should begin now and it should begin in a very proactive way. >> That was fantastic. Um Shua, thank you so much for your time and thank you for joining us today. >> Great. Thank you. >> And that does it for today's episode. Thank you Shua. Thank you to our panelists Norman, Brad, and Jakamo. Thanks to our viewers and listeners, and thanks to our producers. You can subscribe to Security Intelligence wherever podcasts are found so that you never miss an episode. Stay safe out there.