AI vs. Cybersecurity: What's Actually Changed?

00:00 Hi, everyone. So I'm James Duggan, a Solutions Architect at Synack. But firstly, thank you very much for joining here today. So I'm gonna be talking about AI and in particular some of what's changed recently in cybersecurity. Just a little bit about myself. So I joined Synack around four years ago but prior to that I used to work across government, defense, and security where I was helping deliver various secure cyber solutions and

00:30 secure those solutions themselves. And regularly we had to respond to the rapidly changing threats. But particularly now with AI we see that's accelerating again and particularly that need to respond. So just a little bit about what I'm gonna run through in this session. So firstly, who is Synack and what we do. A bit about the changing threat environment with the ever-enhancing AI capability. A little bit about that resulting discovery wave and

01:01 what that means, the coverage gap, and then the importance of the harness and guardrails in offensive security testing. And I'm gonna give an an example of an agentic pen test system and then how you can validate those findings So a little bit about who Synack are. So Synack were founded around twenty thirteen and initially kind of focused on providing managed bug bounty and pen testing for external perimeter and applications. And they were founded by a couple of ex-NSA employees,

01:34 particularly to kind of to solve that diversity and skills challenge with security testing, but also bring in a much more adversarial way of testing. So what's Synack do? They bring fifteen hundred researchers that are incentivized to find vulnerabilities, and that really helps them emulate the threat. And they do so in a way that really helps sort of bring that extra sort of diversity to that testing and with complete control over that testing experience with associated analytics

02:06 helping really show proof of coverage as well. And at Synack there's two, two main models to that testing. Firstly, Synack Red Team, which is our elite vetted researchers, and then secondly, the Sara agentic AI pen test capability. So that's the Synack autonomous RED agent. So that really helps Synack cover that full life sa- cycle through from discovery, testing, vulnerability management and reporting. So I just want to cover a little bit about

02:36 some of the changes in the threat landscape, and this has really changed a lot over the last year or so. So for twenty twenty-six it's all about scale, speed and autonomy. So the first one here the lowered barrier to entry. So years ago, you ne- needed to be a real elite coder to really be able to find exploitable vulnerabilities in systems. And now with AI tools, this can really make a weak attacker

03:06 much more effective than what they might have been previously. Give you an example teenagers with next to no coding background leveraging ChatGPT to attack Rakuten Mobile across about two hundred and twenty thousand times. Second on this list, we've got machine speed exploitation. So we're now seeing over twenty-eight percent of CVEs being exploited within just twenty-four hours of disclosure, and I'll come onto that a little bit more in a moment.

03:37 And then when you couple that with dev teams shipping hourly, that twice a year or traditionally once a year pen test is just no longer good enough. And then agentic autonomy. So previously, we might have had some of the most sophisticated AIs only being able to do a couple of steps within an attack chain, but now that multi-step autonomy, autonomy is becoming much more sophisticated. So when we look at from discovery, validation, through to proof of concept,

04:07 exploit construction, we've got AI agents able to, for example, run full ransomware intrusion in just hours, compared with operating alone, which would've been over a few weeks. And then number four, looking at the AI supply chain itself. We are hearing about AIs going rogue and Hugging Face, for example, being attacked. But what about where you might have threat actors which have compromised the AI models which companies might be using, such as LightLLM or Trivy?

04:38 And this impacted around two and a half thousand organizations, resulting in exposing sensitive information such as keys, passwords, and credentials. So I mentioned about kind of that increased speed and that need to respond much, much faster. So that window between disclosure and exploitation really is collapsing. So the diagram shown on this slide is from the zero-day clock, and this is showing that meantime to exploit, which is falling from what was previously years to now being just, just hours.

05:12 So if we go back to 2018, we were seeing a mean time to exploit of a couple of years, so seven hundred and seventy-one days. And then this has really had a rapid downward trend. So now we're like, wow, quarter one of 2025, about twenty-eight percent of known exploitable vulnerabilities were having exploitation evidence within just one day of that CVE being published. And then moving forward to the first half of this year, seeing twenty-three percent of known exploitable

05:44 vulnerabilities being exploited on or before disclosure days. And then just a stat from IBM X-Force. So seeing a forty-four percent, an increase in attacks that began with exploitation of public-facing applications. And one of the sort of largely driving aspects of that was as well as authentication controls, is believed to be AI-enabled vulnerability discovery. But if we're comparing that to defenders, it hasn't really changed that much.

06:16 So a stat from Verizon for that meantime to remediate- is around 32 days or 38 days if you're looking across the known exploitable vulnerabilities. That 32 days was just for the, the internet-facing ones. And this really does sort of highlight that need to discover and remediate much faster. So being able to patch before exploit- exploitation, though, we have to kind of accept that isn't always realistic to do. So I mentioned before about the multi-step

06:48 autonomy being something that's rapidly increasing. So there is a test put in place by the UK AI Security Institute, and it's a 32-step corporate network which was produced called The Last Ones And this was essentially a test from reconnaissance through to objective. And this was something which was typically taking a human expert somewhere between fourteen and twenty hours. And then a number of various AI models were run against this. I'm particularly looking at frontier models with a fixed inference

07:22 budget. So if we just kind of look at how quickly this has progressed, going back to twenty twenty-four, in August, it was just one point seven steps, nine point eight steps early this year. So that was a six-fold increase in just eighteen months already. And then moving forward just five months later in around July there were now models at which we're able to s- to complete the full attack chain of thirty-two steps in eight out of ten attempts on that same range.

07:52 So I think just really highlighting that the rate of progress on that multi-step autonomy is really rapid. So that doubling time taking less than five months recently. So that multi-step autonomy threat really is becoming real. But one of the things that we're seeing in particular is that the rate of discovery really has grown significantly. So as this title here says, vulnerability's gone machine rate, remediation hasn't.

08:23 So some of the same constraining factors, they still remain. So the constraining factor on triage but then also on va- validation, yet the discovery aspect of it has become relatively easy. So this chart here is just one from Anthropic which was ran for their AI vulnerability discovery program, where they ran it across a thousand plus critical open source projects. And just to kind of read through a few stats on on here. So this is back from around August. So over twenty-six thousand findings discovered,

08:58 and five thousand plus triaged, then four and a half thousand valid, a couple of thousand then reported to maintainers, and four hundred and twenty-one patched upstream. And really just again, just that point to highlight that finding vulnerabilities has become really quick. Discovery has become much easier, but we're really just sort of limited by that time it takes to triage and then remediate. And just to kind of put that into perspective in terms of how some of industry is responding to that.

09:29 So, curl that audited their AI, those reports from AI one by one. So of all of the reports which were submitted there were five reported as confirmed security vulnerabilities, one of which was confirmed as a low-severity CVE, three of which false positives, and these were ones which had already been documented. And we're seeing that many of the providers as a result had been dropping some of their previous bug bounty testing programs due to that

10:01 vulnerability wave, which isn't sustainable to validate those at the re-required rate. For Mozilla, so they had over two hundred and seventy vulnerabilities reported related to Firefox vulnerabilities, and then three of those CVEs were credited to the AI alone within the security advisory, and many of which were determined to be vulnerabilities, so either low severity or bugs which were found in non-exploitable code paths, and many of which, which were determined that

10:32 they could have been found by a human researcher, so it wasn't that AI alone were able to find those. And then just another stat here from VulnCheck, just highlighting that a huge proportion of their vulnerabilities were not necessarily confirmed to be attributable to AI alone and were kind of not necessarily exploitable in the wild. So really just, just wanted to kind of summarize again that there's a big difference between finding or discovery

11:03 of those vulnerability versus actually confirming that they really are exploitable vulnerabilities. And in many cases, a huge amount of money is spent on the inference budgets or in order to find those vulnerabilities or various protections such as running them in sandboxes or other mitigations had been put in place. Then it's a really important factor if someone tells you that they've found thousands of vulnerabilities, firstly, what was the severity of those vulnerabilities?

11:33 Are they actually confirmed exploitable, and who confirmed that they are? So just to kind of call out a stat here, and this is back from November '25, but just kind of put that baseline into on, into perspective. And this is where it was kind of noted then from the Google Threat Intelligence group that Claude was frequently overstating findings, occasionally fabricating data during autonomous operations where it might have claimed to have obtained credentials that then didn't work. And that's a really big obstacle in kind of that fully autonomous cyberattack.

12:06 As mentioned previously, that sort of multi-step autonomy really is increasing rapidly. So really sort of highlights that AI testing alone hallucinations do really remain a particular limiting factor from having sort of purely sort of hands-off cyber operations. And then NCSC recently also mentioned that certainly for now, they do expect humans to remain in the loop. So just talk a little bit about the coverage gap. So this is from a study which Synack commissioned to Omdia,

12:39 and it's highlighted that despite most organizations, so ninety-five percent in this case, really prioritizing pen testing, still a relatively small proportion of that attack surface is tested. So around thirty-two percent of the attack surface is tested. So really just wanna highlight this because particularly when we look at now that sort of discovery wave or the increased ability to find those vulnerabilities, the increased speed in which they're find- being found,

13:10 that coverage gap becomes an even more critical factor that we can begin to close that. So when you look at offensive AI, which can rapidly map that attack surface, identify weak points, iterate on that at machine speed, everything's now becoming a potential attack vector, whether you're looking at, like, aging firewalls or routers or API endpoints or even just it might be applications which haven't been tested at the same cadence as what you're releasing.

13:41 And I think when you're looking at this coupled with now people releasing capabilities at a much frequent cadence with vibe coding as an example, that becomes an even bigger challenge when you're coupling that with that coverage gap. to know your perimeter, not just knowing your scope that you're currently testing So what do we see as some of our recommendations? Certainly some actions that you might wanna

14:13 take at the moment. So a lot of this are things that people might already be doing but particularly need to go faster or at a greater scale. So firstly kind of as related to that coverage gap, increase coverage certainly, certainly where viable for your attack surface in terms of what you're testing. And have a good inventory of what you're testing as well. And also, where possible, increase the frequency at what you're testing. And in some cases, leverage some of the defensive AI

14:44 capabilities for such as agentic pen testing to help you do that at a, at a speed which is now much more viable than what it would've been with relying on humans alone. That really kind of leads to the second point of increase that speed of testing to m- match machine speed But in doing so, with the sort of wave of vulnerabilities that you might be seeing from discovery, make sure you're validating those, and then prioritize to try and minimize that that exposure time.

15:16 So you can both identify information, use the analytics to help you work out what to focus on. Couple, couple that with the inventory information that you might have, with the priority information, with the business context, to really help you shrink that exposure time on those assets which matter most to you and on those vulnerabilities which matter most. So I'm now just gonna move on to an example of an agentic pen test system. So for Sara, the Synack Autonomous Red Agent. So this is an example for the Synack capability.

15:48 We've got an org- We've got our own agentic pen test capability, which we've been developing over a couple of years. We've got human in the loop for this, which essentially means that we will be validating, checking all of those findings, making sure that we are not providing any noise to our customers and that they are truly exploitable vulnerabilities. But in terms of how this works, so firstly, we've got an orchestrator which oversees all of this.

16:18 So th- this looks at, for example, the global state. It's coordinating specialist attack agents. It manages the strategy of those resources. And really thinking of this as the part of the process which keeps all of those individual elements in sync. And then there's five main parts to this. Firstly, there's reconnaissance agents. So this is gonna be running a number of specialist reconnaissance elements which all run in parallel to really understand what it is that's being tested.

16:51 So for example, understanding information about the web applications or the infrastructure, like looking at what ports are open or what technologies are being used for web applications. Based on what's been seen from the reconnaissance agents a number of attack agents are then gonna be then deployed in parallel. So there's essentially gonna be a swarm of attack agents, which are then going to, going to be r- or run a specialized agent based on those findings. And then depending on what's been found we're then

17:23 gonna have independent verification agents, which are gonna be checking those findings to making sure what's being reported is actually exploitable. But then as a final step, as part of our human-in-the-loop element, we've got internal vulnerability operations team, which is gonna be validating those reports and providing the same level of checks to what we would do from our exploitable vulnerability submissions from our Synack Red Team. So this really allows us to be having that same level of quality and

17:54 checks to what we would do for our human testing And that's gonna be making sure that those vulnerability submissions all are actually exploitable and also contain the information they need to contain, like detailed steps to reproduce, recommended fix screenshots where relevant to provide details in order to reproduce those vulnerabilities. And then as part of ensuring that this was done in a really secure, safe way, we've got guardrails that oversee this, and this is gonna be making sure that those attack agents stay within scope,

18:29 that they don't allow any destructive behavior that we, for example, prevent if a vulnerability within a-- getting access to a database such as a SQL injection was found, we prevented the ability to go further, such as dropping tables and behaviors that you wouldn't want to see as a, as a customer. So really just to kind of close it out, so we've got that sort of structured orches- orchestration with that adaptive intelligence.

18:59 And this is really making sure that we can then deliver those results at scale with the audit trails compliance that you'd expect, but in an enterprise way with role-based access control and reporting that you would need to enable you to act on those findings. So just to quickly call out an example from a recent Sara test that we ran for a for a customer. Firstly, just wanna call out the kind of the speed. So this ran in just six hours, so really, really quick.

19:30 And it found three really, really important vulnerabilities to start with and three each important in its own right. So firstly, we were able to uncover account takeover, which was found through a reset token something which was found in the response body, and Sara confirmed essentially they could log in with new credentials. And then a second exploitable vulnerability related to SQL injection, where through reconnaissance of a client-side script not necessarily in the user interface itself,

20:02 but within the kind of client-side script, it was able to identify sort parameter issue whereby email addresses and hash passwords were essentially being exposed And then also a stored cross-site scripting vulnerability whereby a persistent payload was a-- vulnerability was found which could essentially be deployed into any victim's browser enabling further takeovers. So each of these is really kind of important in its own right, but combined they could lead to a complete organizational compromise.

20:36 And each one of these so seventy percent of the findings were high or critical in this case. So yeah, as as we sort of talked about earlier, so discover-discovery is no longer the bottleneck. It really comes down to now that need to kind of validate, but then also put it in a way which enables us to remediate faster to minimize that exposure time. So just a little bit about some of what we're doing on the validation side. So firstly Sara, which is that agentic pentester, flags lots of issues.

21:09 But Sara itself is verifying them through that verification stage within the process before it gets reported to a human. And this is just one example just from a sample of information whereby we flagged about fifteen thousand plus possible issues and then reduced that down rapidly to reporting for human validation of about eight percent of those. And then our internal vulnerability operations team then check what's actually exploitable before we report that to you.

21:39 The right side of here is just showing some examples of some of the different types of exploitable vulnerabilities which were being challenged. And the pattern, pattern-based ones rare-rarely being challenged. It's more some of those complex, more nuanced ones, maybe from SQL injection or maybe from some of the business logic ones. And these are just examples of some of the samples that we took. And this is really sort of highlighting that that again, that sort of validation element,

22:10 that need to kind of check those before they're being presented and potentially leading to ala-alarm bells going off that it has already been confirmed to be exploitable first. And here just kind of sets out some of those processes that we go through within then our human level of validation. Firstly assuring that it is rep-reproducible to checking each of those steps and evidence is, is sufficient. Is it within the scope that was provided? Has it met the s-- that level of acceptance?

22:43 And then does the impact represent what's been found by the internal team? And then all of these vulnerabilities will essentially be reproduced internally before they're then reported. So AI assistance is used to enable us to cope with that increased wave of vulnerabilities, such as finding duplicates, doing some of that proofreading, making sure that it's in scope. But that final human reproducible check is c- is checked before we actually make sure that

23:15 vulnerability is either accepted or rejected So finally, I just want to kind of call out that we're seeing three, three main models emerging in pen testing. Firstly the traditional human pen testing, absolutely great for that sort of creativity and depth, particularly when you've got a crowd of researchers which are bringing that sort of diversity to testing, enabling you to emulate the threat.

23:46 That's sort of purely agentic model, but it does come with risk such as hallucinations and if you're then purely using an AI model to report evidence, that leads to, leads to challenges. And then having that augmented model, so human testing augmented with AI, with humans validating or operating side by side. So certainly we see at the Synack position is to have human testing augmented with AI enabling us to bring the best of both,

24:18 that human testing for that depth and that human creativity that you require, but also AI testing for that additional speed and able to keep pace with the increased threat. So thank you very much for listening. If you've got any questions please get in touch. We'd love to talk to you further. Thank you very much.

No lines match that search.