Latest / Expert Insights Podcast / #95. Black Hat 2026: AI Agents, Frontier Models, and the Quantum Countdown
Transcript
- 0:03SPEAKER_00: Hello and welcome to the Expert Insights Podcast.
- 0:06SPEAKER_00: With Black Hat 2026 having wrapped up last week,
- 0:10SPEAKER_00: two topics dominated every conversation that we had.
- 0:13SPEAKER_00: Agentic AI and how fast it's changing both sides of the fight,
- 0:17SPEAKER_00: and the growing pressure to get ready for a threat landscape
- 0:20SPEAKER_00: that isn't waiting around.
- 0:22SPEAKER_00: In this roundup, we're bringing you four of our favourite conversations from the show floor.
- 0:27SPEAKER_00: We'll hear from Proofpoint on how attackers are compressing their timelines
- 0:31SPEAKER_00: and putting AI to work in live campaigns,
- 0:34SPEAKER_00: from two Zscaler leaders tackling agentic AI from two very different angles,
- 0:39SPEAKER_00: identity risk and hands-on frontier model testing,
- 0:42SPEAKER_00: and we'll close with a conversation on a threat that's easy to lose in all
- 0:46SPEAKER_00: the AI noise, quantum.
- 0:48SPEAKER_00: Let's get into it.
- 0:50SPEAKER_00: First up, Daniel Blackford, VP and Head of Threat Research at Proofpoint.
- 0:55SPEAKER_00: Daniel's team spends its time inside the data, hunting for what's new in adversary behaviour.
- 1:01SPEAKER_00: We started by asking him what's actually changed in the threat landscape this year,
- 1:05SPEAKER_00: and his answer had less to do with new tactics than with speed.
- 1:10SPEAKER_02: A ton of threat actor activity is hosted behind Cloudflare, right?
- 1:12SPEAKER_02: It's like as more and more entities are involved in the same attack chain,
- 1:17SPEAKER_02: overall attribution becomes much more difficult.
- 1:20SPEAKER_02: So like that's the ecosystem we're in now.
- 1:23SPEAKER_02: And then you add...
- 1:26SPEAKER_02: Kind of facilitation of AI-assisted tools on top of that.
- 1:31SPEAKER_02: And that,
- 1:32SPEAKER_02: you know,
- 1:32SPEAKER_02: the trend that everyone needs to be aware of right now is not necessarily
- 1:37SPEAKER_02: a large shift in the tactics themselves.
- 1:41SPEAKER_02: Techniques and procedures, but it's, you know, the rate and scale of execution.
- 1:46SPEAKER_02: It's like very compressed timelines.
- 1:49SPEAKER_02: And so actors, like, they're people.
- 1:53SPEAKER_02: They don't want to spend literally 24-7 hours
- 1:57SPEAKER_02: They're not hacking into things either.
- 1:58SPEAKER_02: They want to make some money or complete their espionage work and use that
- 2:04SPEAKER_02: to go home to their family, their significant others, take some vacations and stuff.
- 2:09SPEAKER_02: And so literally there's like a throughput of how many victims could be processed in
- 2:15SPEAKER_02: a small amount of time.
- 2:17SPEAKER_02: And you could see this even going back to TA-505 when they kind of first...
- 2:25SPEAKER_02: We're processing CLOP ransomware infections, which are still active.
- 2:29SPEAKER_02: And we just saw another,
- 2:30SPEAKER_02: we just saw a very recent campaign from them so that like they're still at it.
- 2:34SPEAKER_02: But you could see the amount of potentially infected victims was far less than
- 2:41SPEAKER_02: the number of ransoms that were actually happening because it takes time
- 2:46SPEAKER_02: to process them,
- 2:47SPEAKER_02: to investigate,
- 2:49SPEAKER_02: you know,
- 2:49SPEAKER_02: the Valuable environment,
- 2:51SPEAKER_02: move laterally through it,
- 2:52SPEAKER_02: and then hands on keyboard actually deploy their ransomware.
- 2:58SPEAKER_02: These timelines,
- 2:59SPEAKER_02: the kind of judging from all of the available potential infections,
- 3:06SPEAKER_02: who is the most lucrative target,
- 3:08SPEAKER_02: and then actually escalating privileges and migrating laterally through
- 3:14SPEAKER_02: that network, doing the kind of network exploration part of it.
- 3:18SPEAKER_02: You're able to do that much faster at a larger scale now.
- 3:21SPEAKER_02: So that, I guess, is the biggest thing.
- 3:27SPEAKER_02: You know,
- 3:28SPEAKER_02: businesses,
- 3:28SPEAKER_02: especially ones that are operating at a large size,
- 3:30SPEAKER_02: may have gotten accustomed to being able to patch on a certain timetable now
- 3:38SPEAKER_02: or they have a certain amount of time to investigate and remediate an incident,
- 3:43SPEAKER_02: right?
- 3:44SPEAKER_02: But the time is the part that's kind of collapsing.
- 3:48SPEAKER_00: That questioning of speed leads directly into how attackers are actually using AI today.
- 3:53SPEAKER_00: And Daniel had specific current examples.
- 3:57SPEAKER_02: Vibe coding is the great example.
- 3:59SPEAKER_02: So, you know,
- 4:00SPEAKER_02: we have seen them Vibe code up a lot
- 4:03SPEAKER_02: of various phishing pages using technology like lovable AI.
- 4:07SPEAKER_02: You can just add a prompt and spin up a beautiful curriculum.
- 4:12SPEAKER_02: Providential harvesting page,
- 4:13SPEAKER_02: same goes for the kind of initial,
- 4:15SPEAKER_02: you know,
- 4:15SPEAKER_02: whatever initial social engineering content is put before an intended victim,
- 4:19SPEAKER_02: you know,
- 4:20SPEAKER_02: could be in the actual email body,
- 4:23SPEAKER_02: it could be HTML,
- 4:26SPEAKER_02: it could be a different type of landing page, like that part is definitely obvious.
- 4:33SPEAKER_02: But we're seeing them kind of start to pivot now.
- 4:37SPEAKER_02: I think there's like a huge question of,
- 4:40SPEAKER_02: well, are we going to see these systems attacked directly?
- 4:45SPEAKER_02: You have to consider the fact that agents acting semi-autonomously with some amount
- 4:51SPEAKER_02: of permissions relative to the data, right?
- 4:53SPEAKER_02: Like this is a huge risk.
- 4:54SPEAKER_02: I think businesses recognize that.
- 4:56SPEAKER_02: And maybe
- 4:59SPEAKER_02: this time, as opposed to Mass cloud adoption.
- 5:04SPEAKER_02: They're taking it very seriously and figuring out how to secure it ahead of time.
- 5:10SPEAKER_02: But of course, threat actors will experiment.
- 5:15SPEAKER_02: And so while we're not necessarily seeing,
- 5:17SPEAKER_02: for example,
- 5:19SPEAKER_02: prompt injection injected into email content at scale, there's some of it coming in.
- 5:25SPEAKER_02: And then recently we just published some research where From an underground forum,
- 5:29SPEAKER_02: we've started to see threat actors,
- 5:31SPEAKER_02: to your point, the kind of as-a-service model, advertise indirect prompt injection.
- 5:37SPEAKER_02: So this is, you know, passively inserting text into information.
- 5:43SPEAKER_02: I mean, there are tools now
- 6:01SPEAKER_02: I think it's very much a situation of if it's not broke, you don't need to fix it.
- 6:09SPEAKER_02: You can look at the light LLM.
- 6:16SPEAKER_02: Intrusion, it wasn't an abuse of AI technologies that allowed that to happen.
- 6:23SPEAKER_02: It was credential theft.
- 6:26SPEAKER_02: So those things aren't necessarily going away.
- 6:28SPEAKER_02: But we see actors building their tools.
- 6:35SPEAKER_00: One story from Daniel's team stood out above the rest this year.
- 6:38SPEAKER_00: A Russian state-sponsored vulnerability proof point has nicknamed Half Click.
- 6:43SPEAKER_02: We're calling it half-click because,
- 6:47SPEAKER_02: to your point,
- 6:49SPEAKER_02: the vulnerability ran code simply from the mail being loaded in the preview frame,
- 6:57SPEAKER_02: and so it didn't require And also, we say no social engineering.
- 7:00SPEAKER_02: I guess you could probably
- 7:13SPEAKER_02: Yeah.
- 7:17SPEAKER_02: It's like,
- 7:17SPEAKER_02: how granular do we want to get with the psychology of convincing some,
- 7:21SPEAKER_02: you know, some people open everything.
- 7:23SPEAKER_02: Some people are probably deleting things without ever looking at it.
- 7:27SPEAKER_02: But it is.
- 7:28SPEAKER_02: It's always more dangerous,
- 7:30SPEAKER_02: I think,
- 7:31SPEAKER_02: when you don't have to convince someone to do something that they're trained not
- 7:37SPEAKER_02: to do.
- 7:38SPEAKER_02: And so, yeah, in this particular case, it's not just...
- 7:44SPEAKER_02: It's not just this particular Russian group and Zimbra.
- 7:51SPEAKER_02: We published three different articles about various Russian nexus activities.
- 7:59SPEAKER_02: There's a lot of activity now at this point on different mail servers, right,
- 8:03SPEAKER_02: that are all kind of processing things in a way,
- 8:08SPEAKER_02: processing the HTML in a way that is going to allow for,
- 8:12SPEAKER_02: you know, remote code execution.
- 8:15SPEAKER_02: RoundCube was another one that we saw both Russia and China recently exploiting.
- 8:22SPEAKER_02: Now that it's out there,
- 8:24SPEAKER_02: maybe even services that have not explicitly been exploited,
- 8:27SPEAKER_02: and we've worked directly with these vendors to notify them of the vulnerability
- 8:31SPEAKER_02: before we ever went public.
- 8:35SPEAKER_02: It gives them an opportunity to maybe fix their own issues ahead of time as part
- 8:41SPEAKER_02: of why we're doing this,
- 8:42SPEAKER_02: not just to frustrate the adversary, but also highlight opportunities
- 8:50SPEAKER_02: We didn't want to end up on a purely defensive note,
- 8:52SPEAKER_02: so we asked Daniel about the work being done to actually take these operations down.
- 9:17SPEAKER_02: Operation Endgame is ongoing.
- 9:21SPEAKER_02: We're in phase four now.
- 9:25SPEAKER_02: And I think,
- 9:27SPEAKER_02: you know,
- 9:27SPEAKER_02: the theory behind it in general and part of why it's been so successful is
- 9:32SPEAKER_02: because it did stop focusing on trying to fight ransomware and instead focuses
- 9:39SPEAKER_02: on fighting the ecosystem because ransomware...
- 9:42SPEAKER_02: Generally, the actual event is not going to happen unless there's some initial access provided.
- 9:50SPEAKER_02: And these actors are not all the same.
- 9:52SPEAKER_02: There are a lot of different initial access brokers.
- 9:55SPEAKER_02: But if you can disrupt the distribution system and what's getting the kind
- 9:59SPEAKER_02: of landing on the beachhead that's ultimately going to lead to the ransomware,
- 10:02SPEAKER_02: then you're functionally disrupting the ransomware.
- 10:04SPEAKER_02: And so treating it as an ecosystem and looking at...
- 10:08SPEAKER_02: Money mules and payment processors and crypto tumblers and finding who is connected
- 10:14SPEAKER_02: to what and how can we disrupt any of these elements as opposed to just kind
- 10:22SPEAKER_02: of laser focusing on one particular actor group is having a lot of success, right?
- 10:27SPEAKER_02: Because, you know, we obviously have a lot of data.
- 10:30SPEAKER_02: We see a number of these actors,
- 10:34SPEAKER_02: definitely ones that We're using a lot of the major botnets
- 10:35SPEAKER_02: that have been disrupted, TrickBot, Cubot, iCity, kind of all of these, right?
- 10:41SPEAKER_02: They were distributing campaigns at scale,
- 10:45SPEAKER_02: anywhere from,
- 10:46SPEAKER_02: you know,
- 10:47SPEAKER_02: 20,000 to 2 million emails in a given day,
- 10:50SPEAKER_02: which we kind of put, you know, wrap up, we call it a campaign, right?
- 10:55SPEAKER_02: Like a time-bound series of discrete related threat activity.
- 11:00SPEAKER_02: And...
- 11:02SPEAKER_02: You know, we track these things over time in perpetuity, right?
- 11:05SPEAKER_02: And so with a lot of these families that were disrupted, you can see a clear drop off.
- 11:10SPEAKER_02: And in some cases, arrests were made.
- 11:12SPEAKER_02: In some cases, arrests weren't made.
- 11:13SPEAKER_02: Regardless, we've kind of forced a pivot.
- 11:16SPEAKER_02: And And this is kind of at the same time where you see the pivot
- 11:20SPEAKER_02: to things like remote monitoring and management tools.
- 11:23SPEAKER_02: But kind of going back to the philosophy of it,
- 11:27SPEAKER_02: I think my job at Proofpoint,
- 11:30SPEAKER_02: Proofpoint as a company,
- 11:32SPEAKER_02: all of the vendors down here,
- 11:33SPEAKER_02: yes, like our goal is to protect people and our goal is to protect our customers.
- 11:37SPEAKER_02: But I think we have an obligation to In the same way people should be adopting LLMs
- 11:46SPEAKER_02: for their particular use cases right now,
- 11:50SPEAKER_02: we have to use all of the tools at our disposal to make a dent in
- 11:55SPEAKER_02: that $50 billion being handed over.
- 11:58SPEAKER_02: And so when we partner with federal law enforcement,
- 12:02SPEAKER_02: international law enforcement, Intelligence agencies and other government agencies.
- 12:09SPEAKER_02: We're doing what we can beyond our simple customer base to impact that ecosystem.
- 12:14SPEAKER_02: And of course, downstream, it helps our customers, but it helps everyone else too.
- 12:20SPEAKER_02: I think what we have proven is that there is no silver bullet.
- 12:26SPEAKER_02: It takes more than a village.
- 12:28SPEAKER_02: It takes a community.
- 12:30SPEAKER_02: It takes an international cooperative.
- 12:33SPEAKER_00: At Black Hat,
- 12:34SPEAKER_00: Zscaler's Dahul Sharma joined us to talk about a problem that's only getting bigger
- 12:38SPEAKER_00: for security teams,
- 12:40SPEAKER_00: securing the agents themselves, not just the people who deploy them.
- 12:44SPEAKER_00: Here's why he thinks agent identity is a fundamentally different problem from human identity.
- 12:50SPEAKER_01: The AI adoption is going mainstream, especially the agentic adoption in the last few months.
- 12:56SPEAKER_01: So far, the agentic workloads we were seeing were primarily embedded inside SaaS
- 13:00SPEAKER_01: applications like co-pilots of the world and what's running in AWS and stuff.
- 13:05SPEAKER_01: But now people are building their own agents.
- 13:08SPEAKER_01: Agent identities are sprawling and connecting to more and more things,
- 13:11SPEAKER_01: which is creating a little bit of a crisis situation in enterprises
- 13:17SPEAKER_01: where they don't know What identities have access to in terms of identity
- 13:23SPEAKER_01: and data lineage and what actions they can take that can perform any sort
- 13:27SPEAKER_01: of undue incidents.
- 13:32SPEAKER_01: Like everyone probably is talking about the hugging face issue,
- 13:35SPEAKER_01: but if you look under the surface,
- 13:38SPEAKER_01: the problem is that You lowered the guardrails on your models for efficacy
- 13:46SPEAKER_01: and agents to test with that.
- 13:48SPEAKER_01: And then these models basically exploited vulnerability and the gateway controls
- 13:53SPEAKER_01: and went to internet and attacked on other institution.
- 13:59SPEAKER_01: Now, if this would have been a bank,
- 14:01SPEAKER_01: there would have been severe penalties because this is,
- 14:04SPEAKER_01: even if it's a Company,
- 14:07SPEAKER_01: Frontier Lab coming with the right intent,
- 14:09SPEAKER_01: but you compromise a critical asset, you are liable, similar to a threat actor.
- 14:15SPEAKER_01: So this can happen everywhere.
- 14:16SPEAKER_01: That's why need for better governance structure,
- 14:20SPEAKER_01: the need for zero trust architecture in these areas is growing more and more.
- 14:27SPEAKER_01: And what we are seeing is that organizations are Not just talking
- 14:33SPEAKER_01: or evaluating AI security platforms,
- 14:36SPEAKER_01: but also deploying them at a rapid pace as well.
- 14:40SPEAKER_01: When it comes to managing access for agents,
- 14:45SPEAKER_01: how is it different to the kind of human identity and access management
- 14:49SPEAKER_01: that we've seen previously?
- 14:51SPEAKER_01: What are the kind of new challenges there when it is agentic personas that you're controlling?
- 14:56SPEAKER_01: So, first of all,
- 14:58SPEAKER_01: agents either are tied to a human identity or they're autonomous when they're tied
- 15:04SPEAKER_01: to a service principal identity.
- 15:06SPEAKER_01: In either cases, agents have too many permissions.
- 15:09SPEAKER_01: Like your agent that you're setting up is inheriting your authorization
- 15:14SPEAKER_01: and should not be getting access to all things that you can do as a user.
- 15:20SPEAKER_01: It might be created for a task or it can create a certain subagent that is doing
- 15:24SPEAKER_01: a task and if it goes to do that task with the full authorization model of the user
- 15:29SPEAKER_01: and
- 15:31SPEAKER_01: or does things which it should not do, then you as a user are liable.
- 15:37SPEAKER_01: You might be a compromised identity that you need to ascertain
- 15:42SPEAKER_01: if it's getting attached to now agent identity.
- 15:45SPEAKER_01: So you You need to understand that context is very, very important.
- 15:49SPEAKER_01: Agent authentication doesn't solve that problem.
- 15:51SPEAKER_01: You need to build file and gain authorization.
- 15:53SPEAKER_01: This is why it's very important to have visibility not just of
- 15:56SPEAKER_01: where your AI assets are deployed,
- 15:58SPEAKER_01: but how do they connect into your enterprise environment.
- 16:01SPEAKER_01: Ecosystem of data's identities, applications, and so on.
- 16:06SPEAKER_01: Then inspecting the agent traffic once you have basically built
- 16:12SPEAKER_01: that foundational visibility,
- 16:14SPEAKER_01: it needs more than how you have Secured user traffic with firewalls and VPNs
- 16:19SPEAKER_01: and inline proxies is that you need to know the task and intent of the agent
- 16:26SPEAKER_01: and you need to build, keep the memory context, which is expensive.
- 16:30SPEAKER_01: And then being able to apply policy style to that.
- 16:34SPEAKER_01: And then apart from the cybersecurity risk perspective, there is also cost implications of that.
- 16:43SPEAKER_01: Your models using very Expensive,
- 16:48SPEAKER_01: or your agents using very expensive models or actions
- 16:52SPEAKER_01: for very basic tasks could be expensive.
- 16:54SPEAKER_01: So being able to do intent-based routing or LLM routing becomes important.
- 17:00SPEAKER_01: So there are newer set of controls that are required for securing agents.
- 17:05SPEAKER_01: And that's something which you need to build.
- 17:09SPEAKER_01: And also policies for agents,
- 17:10SPEAKER_01: while everyone is still Talking about if you go to OpenAI Anthropic,
- 17:14SPEAKER_01: they order their compliance APIs so people can build controls like DLP,
- 17:18SPEAKER_01: but what you need to build around agentic traffic is not DLP,
- 17:21SPEAKER_01: but more intent-based policies.
- 17:24SPEAKER_01: Why agent was set up?
- 17:26SPEAKER_01: Is it doing the right thing?
- 17:28SPEAKER_01: It might not be violating a DLP policy by,
- 17:32SPEAKER_01: for example,
- 17:32SPEAKER_01: when a support bot agent or you're on
- 17:35SPEAKER_01: a banking website starts giving stock recommendations,
- 17:38SPEAKER_01: but you need to understand with what intent that bot was set up
- 17:41SPEAKER_01: and what it is doing.
- 17:42SPEAKER_01: So being able to build that policy is which needs...
- 17:45SPEAKER_01: Our third guest also came from Zscaler, but from a completely different angle.
- 18:06SPEAKER_00: Through Project Glasswing,
- 18:08SPEAKER_00: an early access program with Frontier Labs,
- 18:10SPEAKER_00: Deepend Desai's team has spent the last 90 days testing models like GPT 5.5 Cyber
- 18:17SPEAKER_00: and Mythos directly against their own infrastructure.
- 18:21SPEAKER_00: We asked him what surprised him most.
- 18:24SPEAKER_01: We were early members of both of these programs.
- 18:28SPEAKER_01: The first 40 members, for instance, in case of Project Glasswing.
- 18:34SPEAKER_01: These models are step function increase in their ability to do multi-step reasoning
- 18:39SPEAKER_01: across a variety of cyber capabilities.
- 18:43SPEAKER_01: So that's number one thing.
- 18:45SPEAKER_01: It's not hype.
- 18:47SPEAKER_01: It is actually a step function change from what the previous models are able to do.
- 18:51SPEAKER_01: And the magic is because of the post-training data.
- 18:54SPEAKER_01: Now, one thing I would like to clarify,
- 18:56SPEAKER_01: because I've heard this from Multiple CXOs who then got access to Mythos
- 19:01SPEAKER_01: and they tried it out and they're like, what are you talking about?
- 19:04SPEAKER_01: We're not seeing same level of effectiveness in things that it finds.
- 19:10SPEAKER_01: The harness that every organization will build around these AI models
- 19:14SPEAKER_01: to perform these,
- 19:15SPEAKER_01: whether it's vulnerability scan,
- 19:16SPEAKER_01: mitigation workflows, it plays a very important role.
- 19:20SPEAKER_01: It is a force multiplier.
- 19:22SPEAKER_00: That distinction between the raw model and the harness built around it turned out
- 19:26SPEAKER_00: to matter for one finding in particular.
- 19:28SPEAKER_01: You run mythos using the hardness that I just described.
- 19:32SPEAKER_01: It will find an issue.
- 19:34SPEAKER_01: It will also generate an exploit code that you can use to prove that this issue is real.
- 19:41SPEAKER_01: It will also generate a patch or a mitigation that you can then apply to fix the issue.
- 19:48SPEAKER_01: When you apply that patch and rerun mythos on it,
- 19:51SPEAKER_01: the non-deterministic nature means Three out of five times,
- 19:56SPEAKER_01: it was able to bypass its own patch and it will find another version.
- 20:00SPEAKER_01: Wow.
- 20:01SPEAKER_00: That kind of unpredictability changed how Deepend's own engineering team works day to day.
- 20:07SPEAKER_01: So we have to do this iteratively two,
- 20:10SPEAKER_01: three times before the patch itself becomes resilient to that type of issue.
- 20:15SPEAKER_01: And we went through that cycle with our infrastructure, our code base.
- 20:20SPEAKER_01: To the point that we now have our engineers running this harness as part of
- 20:25SPEAKER_01: the code check-in process because otherwise
- 20:28SPEAKER_01: the product security function starts becoming the bottleneck.
- 20:32SPEAKER_01: So a lot of this stuff is truly being handled in the shift-left manner.
- 20:37SPEAKER_01: And then the more integrated testing,
- 20:39SPEAKER_01: runtime testing is where product security comes in and do that additional vetting
- 20:42SPEAKER_01: and make sure we handle anything that is missed.
- 20:48SPEAKER_01: That second learning is very important because these are non-deterministic
- 20:52SPEAKER_01: in nature models and they will evade their own patches.
- 20:57SPEAKER_01: So you need to do this iteratively to get to a resilient state.
- 21:02SPEAKER_01: And my guidance to all the CXOs that I've spoken over the last three,
- 21:08SPEAKER_01: four months is even if you don't have access to METOs,
- 21:12SPEAKER_01: It's very important that you start building this muscle around STLC
- 21:16SPEAKER_01: where you start leveraging whichever model is available with that harness.
- 21:20SPEAKER_01: Have your engineering group also start using it and use the patch that the model generates.
- 21:27SPEAKER_01: Yes, it should be human-guetted,
- 21:29SPEAKER_01: but it will speed up the whole process and you will be in a better posture
- 21:34SPEAKER_01: when you get access to more powerful models using that same workflow.
- 21:38SPEAKER_01: Yeah.
- 21:38SPEAKER_01: That's so interesting about it, generating a patch and then working around that.
- 21:43SPEAKER_01: Yeah, I wasn't aware.
- 21:44SPEAKER_01: That's super interesting.
- 21:45SPEAKER_01: Yeah.
- 21:46SPEAKER_01: But have you had any models escape your bedrock environment?
- 21:51SPEAKER_01: That was interesting, right?
- 21:52SPEAKER_01: The most recent story.
- 21:54SPEAKER_01: Yeah.
- 21:56SPEAKER_01: Fortunately, not.
- 21:59SPEAKER_01: And we wanted to make sure the environment in which we have our code repository
- 22:03SPEAKER_01: and things does not have open internet access.
- 22:06SPEAKER_01: So it was fairly contained.
- 22:08SPEAKER_01: Yeah.
- 22:10SPEAKER_01: I would say the third finding was the whole chaining piece,
- 22:14SPEAKER_01: like ability to chain multiple issues.
- 22:17SPEAKER_01: We've all been talking about it theoretically for the last several months.
- 22:22SPEAKER_01: Seeing it in action and the ability to do it in minutes versus days was very, very interesting.
- 22:29SPEAKER_00: Of everything his team tested against these models,
- 22:32SPEAKER_00: one much older piece of technology held up better than expected.
- 22:36SPEAKER_00: Deepan also had a pointed example of why the relationship between security vendors
- 22:41SPEAKER_00: and frontier labs themselves still needs work.
- 22:44SPEAKER_01: Deception
- 22:45SPEAKER_01: has been a decade-old technology,
- 22:49SPEAKER_01: but I see it becoming a I would say the lowest hanging fruit that you can invest
- 22:55SPEAKER_01: in today in order to get ready to defend your environment against AI-driven attacks.
- 23:00SPEAKER_01: So a common question I got asked again,
- 23:03SPEAKER_01: when you're using these powerful models in your environment,
- 23:06SPEAKER_01: which control did very well in kind of helping you defend?
- 23:11SPEAKER_01: Deception was that control.
- 23:13SPEAKER_01: Inadvertently,
- 23:14SPEAKER_01: these models,
- 23:15SPEAKER_01: because of their ability to chain and find different paths,
- 23:18SPEAKER_01: were tripping on the decoys that were dynamic in nature.
- 23:23SPEAKER_01: As soon as that decoy gets triggered, it will get destroyed and another one will get spun up.
- 23:29SPEAKER_01: And you will have multiple opportunities to flag these attacks.
- 23:33SPEAKER_01: Now, deception is a standalone technology.
- 23:37SPEAKER_01: You are still reacting.
- 23:39SPEAKER_01: Meaning,
- 23:39SPEAKER_01: even if an AI tripped a decoy,
- 23:42SPEAKER_01: they will continue,
- 23:43SPEAKER_01: they will eventually find out this is a decoy
- 23:45SPEAKER_01: and they will continue going down other paths to the crown jewel.
- 23:48SPEAKER_01: Yeah.
- 23:49SPEAKER_01: So what becomes important is as soon as a decoy gets tripped,
- 23:53SPEAKER_01: Deception should be part of your Zero Trust stack,
- 23:56SPEAKER_01: which allows you to take action at the time that attack is happening,
- 24:00SPEAKER_01: which is where the difference is.
- 24:01SPEAKER_01: So, allowing you to respond at machine speed using deception is where you're able to
- 24:06SPEAKER_01: toward these AI-driven attacks.
- 24:08SPEAKER_00: We'll close this roundup with a different kind of urgency.
- 24:11SPEAKER_00: While most of Black Hat was talking about AI,
- 24:14SPEAKER_00: Todd Moore,
- 24:15SPEAKER_00: global VP and GM of data security products at Talis,
- 24:19SPEAKER_00: wanted to make sure Quantum didn't get lost in the noise.
- 24:22SPEAKER_00: We asked him why security leaders needed to be thinking about this now, not in a few years.
- 24:29SPEAKER_03: We'll see you next time.
- 24:44SPEAKER_03: You know,
- 24:44SPEAKER_03: a quantum event that impacts our current cryptography around the 2029,
- 24:49SPEAKER_03: 2030, if you listen to Gartner and Google and a few other large organizations.
- 24:54SPEAKER_03: With regulations coming from the US government,
- 24:57SPEAKER_03: from France recently,
- 24:58SPEAKER_03: coming out of Singapore,
- 24:59SPEAKER_03: Australia,
- 25:00SPEAKER_03: a lot of governments are now being prepared for this advent and asking
- 25:05SPEAKER_03: that companies start implementing safeguards into their systems.
- 25:08SPEAKER_03: And like any large cryptography, you know, cryptography underpins everything you do, right?
- 25:14SPEAKER_03: Our lifestyle, our banking, how we interact with the internet.
- 25:17SPEAKER_03: And any sort of large rotation of algorithms that we've done in the past,
- 25:21SPEAKER_03: whether it was going to AES or going from SHA-2 to SHA-3, it's not trivial.
- 25:25SPEAKER_03: It's not trivial because photography is pervasive within our systems.
- 25:28SPEAKER_03: And so I think the risks are becoming more relevant.
- 25:31SPEAKER_03: There's governance that says we have to start moving.
- 25:34SPEAKER_03: And we're starting to see industries really getting serious about,
- 25:37SPEAKER_03: my gosh, you know, 2029 is not that far away.
- 25:39SPEAKER_03: Think about it.
- 25:39SPEAKER_03: We're almost out of 2026.
- 25:41SPEAKER_03: You're talking two years.
- 25:42SPEAKER_03: So it's now really a relevant time to start getting prepared.
- 25:46SPEAKER_01: And these processes take a lot of time to put them back.
- 25:48SPEAKER_01: They do, they do.
- 25:49SPEAKER_01: You can't flick a light switch.
- 25:51SPEAKER_03: And I think that's one of the fallacies.
- 25:53SPEAKER_03: I talk to board level,
- 25:55SPEAKER_03: CISO level folks,
- 25:57SPEAKER_03: and some are not in the regulated industries, but in broader industries.
- 26:01SPEAKER_03: Who will be impacted too, saying, well, can I just, you know, flip a switch?
- 26:04SPEAKER_03: Can I just implement these new algorithms and it's just going to work?
- 26:07SPEAKER_03: And it's not the, you know, as we've seen in the past, it doesn't work that way.
- 26:10SPEAKER_03: With PQC cryptography,
- 26:13SPEAKER_03: the new algorithms that are coming out that have been ratified by NIST and others,
- 26:16SPEAKER_03: different key lengths, different block sizes, different latency.
- 26:20SPEAKER_03: It's going to really...
- 26:22SPEAKER_03: You know, change your network and impact your infrastructure.
- 26:24SPEAKER_03: So you can turn on these new algorithms,
- 26:26SPEAKER_03: but if you don't prepare your applications,
- 26:28SPEAKER_03: prepare your infrastructure, you're going to break stuff.
- 26:31SPEAKER_03: So it really is, it has to be a process and a plan to get prepared.
- 26:34SPEAKER_01: Absolutely.
- 26:35SPEAKER_01: And we've heard a lot about AI this week.
- 26:37SPEAKER_01: How do you see the rapid adoption of AI changing the data security picture?
- 26:41SPEAKER_01: And do you think the AI challenge and the quantum challenge is becoming linked?
- 26:45SPEAKER_03: Absolutely.
- 26:46SPEAKER_03: I think it's one of the messages here,
- 26:47SPEAKER_03: as you know,
- 26:48SPEAKER_03: on the on the showroom floor and maybe in your conversations, it's dominated by AI.
- 26:52SPEAKER_03: I mean, artificial intelligence,
- 26:53SPEAKER_03: but you know,
- 26:54SPEAKER_03: all those agents,
- 26:55SPEAKER_03: they have personalities, they've got credentials, they got identities.
- 26:59SPEAKER_03: They're talking to your data and they're doing it using security and
- 27:02SPEAKER_03: that security is based So I think it's got to be...
- 27:30SPEAKER_03: You know, known and we want to bring to the surface that AI really depends on these methods.
- 27:35SPEAKER_03: These methods are changing.
- 27:36SPEAKER_03: So AI needs to be part of the PQC and the quantum discussion.
- 27:40SPEAKER_03: Now, one of the good things is, Joel, is that AI actually can be used to help you.
- 27:45SPEAKER_03: I think that's what people forget,
- 27:46SPEAKER_03: that you can actually use AI to help you find where there's gaps in your systems,
- 27:50SPEAKER_03: and you can actually start preparing for this quantum new world with the use of AI.
- 27:56SPEAKER_03: And the last point I'll make is that when I talk to people here on the showroom
- 27:59SPEAKER_03: for everyone, Everyone says, I have so much budget for AI, how do I use it?
- 28:04SPEAKER_03: I would argue that when you talk about PQC and getting ready for quantum with some
- 28:08SPEAKER_03: of those leaders, they'll tell you I have no budget for quantum.
- 28:12SPEAKER_03: It'd be great if we can find a way to convince our leaders what we just talked about.
- 28:16SPEAKER_03: AI security relies on quantum.
- 28:18SPEAKER_03: Give me some of the AI budget to actually go off and do something to protect against quantum.
- 28:21SPEAKER_03: Hope that makes sense.
- 28:23SPEAKER_03: I do think it's all coming together.
- 28:25SPEAKER_01: That's really, really interesting.
- 28:26SPEAKER_01: And I think it's good for...
- 28:28SPEAKER_01: It turns out the AI conversation and the quantum conversation are more connected
- 28:33SPEAKER_01: than most people realise.
- 28:44SPEAKER_03: I do think the first step is try to do your best to get an inventory.
- 28:47SPEAKER_03: Understand what applications,
- 28:50SPEAKER_03: what systems,
- 28:51SPEAKER_03: what infrastructure is leveraging this different legacy cryptography,
- 28:55SPEAKER_03: RSA and ECDSA, that's going to have to change.
- 28:58SPEAKER_03: You may not be able And then next,
- 29:00SPEAKER_03: pick a couple applications,
- 29:00SPEAKER_03: some really key applications,
- 29:02SPEAKER_03: core applications for you as a business, and really start testing what
- 29:14SPEAKER_03: What it doesn't mean to turn on these new algorithms
- 29:17SPEAKER_03: and these new quantum safe mechanisms,
- 29:19SPEAKER_03: these new protocols,
- 29:20SPEAKER_03: and see what it does to your system, what it does to your applications.
- 29:22SPEAKER_03: And so it's going to be a phased approach.
- 29:24SPEAKER_03: Like any, again,
- 29:25SPEAKER_03: cryptographic standard implementation is going to take some time,
- 29:28SPEAKER_03: but I think organizations You need to take an inventory,
- 29:30SPEAKER_03: just a quick inventory,
- 29:31SPEAKER_03: do a risk assessment around that,
- 29:33SPEAKER_03: pick your applications that might be impacted first,
- 29:35SPEAKER_03: the hardest,
- 29:36SPEAKER_03: and then actually do some testing and start slowly working your way
- 29:40SPEAKER_03: through implementing and providing safeguards for all your applications.
- 29:46SPEAKER_01: Todd, you're making it sound so sort of straightforward.
- 29:49SPEAKER_01: It makes me want to go and get
- 29:50SPEAKER_03: started today.
- 29:52SPEAKER_03: But that's just it.
- 29:52SPEAKER_03: I mean, if you're going to be the ostrich with your head in the sand and pretend
- 29:56SPEAKER_03: that it's not going to be your problem for three years, I mean, it takes you time.
- 30:00SPEAKER_03: It's a marathon.
- 30:01SPEAKER_03: Maybe I'll use that analogy, that metaphor.
- 30:03SPEAKER_03: It's a marathon.
- 30:03SPEAKER_03: It's not a sprint.
- 30:04SPEAKER_03: So if you're going to do a marathon, start training now.
- 30:07SPEAKER_03: Start doing little pieces now and you'll get the 26 miles,
- 30:11SPEAKER_03: the marathon eventually, but you got to get started.
- 30:13SPEAKER_03: Yeah.
- 30:14SPEAKER_00: So what should a security team actually do about it?
- 30:17SPEAKER_00: Starting Monday morning,
- 30:19SPEAKER_00: we asked Todd to leave listeners with his top takeaways from the conversation.
- 30:33SPEAKER_03: To starting now, you need to start.
- 30:35SPEAKER_03: And that could be a simple visibility, crypto inventory.
- 30:39SPEAKER_03: It could be a simple risk assessment, looking at applications.
- 30:42SPEAKER_03: Pick one,
- 30:43SPEAKER_03: pick a priority application and start understanding how it would be impacted if
- 30:46SPEAKER_03: a quantum computer popped up.
- 30:47SPEAKER_03: So I think that's...
- 30:48SPEAKER_03: The first takeaway.
- 30:49SPEAKER_03: The second takeaway is watch the hype.
- 30:53SPEAKER_03: There's not going to be a one-all end-all solution.
- 30:55SPEAKER_03: There's not going to be a magic tool.
- 30:57SPEAKER_03: You know,
- 30:57SPEAKER_03: someone that tells you that they have a solution for post-quantum today
- 31:01SPEAKER_03: that isn't crypto agile,
- 31:02SPEAKER_03: both in the algorithm and in the,
- 31:04SPEAKER_03: as we spoke about,
- 31:05SPEAKER_03: the different types of protocols and things like that is not,
- 31:10SPEAKER_03: is not Telling you all the truth.
- 31:13SPEAKER_03: So make sure you're looking and working with vendors that give you the scalability
- 31:16SPEAKER_03: to allow you to grow as the industry changes and grows going forward.
- 31:21SPEAKER_03: You said two or three.
- 31:22SPEAKER_03: What's my third one then?
- 31:24SPEAKER_03: My third takeaway is think about how AI can help you when it comes in solving this journey.
- 31:30SPEAKER_03: I think if you can tie together AI and getting prepared for
- 31:33SPEAKER_03: the first two things we spoke about,
- 31:35SPEAKER_03: I think that's going to help get management buy-in to actually start preparing
- 31:38SPEAKER_03: for this new world.
- 31:39SPEAKER_00: And that's our Black Hat 2026 roundup.
- 31:42SPEAKER_00: Compressed attack timelines and ecosystem takedowns from Proofpoint,
- 31:47SPEAKER_00: agent identity risks from Zscaler,
- 31:49SPEAKER_00: hard evidence from frontier model testing on the ground,
- 31:52SPEAKER_00: and a reminder that the quantum clock is already running.
- 31:55SPEAKER_00: Four different guests, one clear theme.
- 31:58SPEAKER_00: The pace of change is the story, and standing still isn't an option for defenders.
- 32:03SPEAKER_00: If you enjoyed this episode,
- 32:05SPEAKER_00: follow the Expert Insights Podcast wherever you listen and head
- 32:08SPEAKER_00: to expertinsights.com for more coverage from the show floor.
- 32:11SPEAKER_00: Thanks for listening and we'll see you next time.