DSS

DSS 2026 / Day 2 / Pearl Hall

Sword vs. Shield: Navigating the AI Security Chaos in 2027

Davit Kobakhidze

15:47 · Recording

Watch on YouTube ↗Open recording viewerOriginal programme entry
Download plain textDownload timestamped JSON

Source: YouTube automatic caption track · Exported 2026-11-16. Timestamps link to the source video. Cue end times were not supplied.

Transcript

0:00So okay.

0:02Hey, my name is David.

0:04I'm a BD guy from Consensus Diligence.

0:06So not a security auditor.

0:08Sorry about that, but I do think that I

0:11can share some interesting perspective

0:14because I talk a lot with our clients

0:16and

0:17with our security researchers, which can

0:19help me to provide some interesting

0:21information for you guys.

0:22So the the topic of my presentation is

0:25Sword versus Shield. It's an ancient

0:27concept about the competition between

0:29offense and defense team. I believe

0:31originated in

0:33China

0:34thousand years ago. So I do think that

0:37what we have today

0:38on the market it

0:40this competition, who will outsmart who?

0:44The topic says navigating the AI

0:46security chaos in 2027, but also it

0:49backs down to how do you keep your

0:51protocol and funds secure in 2027?

0:55And honestly

0:57after last couple of weeks with all

0:59these bugs exploited, with all the

1:01protocols got hacked,

1:02with all the incident we have

1:05I think you already have an answer. Do

1:08you?

1:09Anyone?

1:11You simply can't.

1:14That's it. That's it for my talk. Thank

1:16you so much.

1:18Okay.

1:20The timer says I still have 14 minutes

1:22to go.

1:23So I perhaps should stop with bombing

1:25with dead jokes.

1:27And let's try to at least figure out how

1:29can we minimize exposure

1:31moving forward.

1:32Before that, let's just have a quick

1:34recap

1:35on what I've heard from the market

1:38throughout this last couple of years.

1:41So first

1:42actually in 2024 I already start hearing

1:45from protocols and security researchers

1:47themselves that oh

1:49AI promises like a lot of efficiency. I

1:52do think it will replace security

1:54researchers moving forward. Back then it

1:56was like

1:57vague timeline three to five years.

2:00But as early as 2026, actually during

2:03the Denver conference, I said here in a

2:05lot from different protocols that you

2:08know what?

2:09We think AI made security simple. We'll

2:12probably cut off our spending and do a

2:15lot of stuff in house.

2:16And throughout 2026 we see this idea

2:19manifested in actions.

2:21So you see like job market for security

2:23researchers in US dropped by 26%.

2:26Lots of security companies, both Web two

2:29and Web three start cutting off their

2:31security teams, auditing teams.

2:34And bug bounties have been paused or

2:36canceled or closed.

2:38But

2:39actually I do think that in reality

2:42right now in November 2026

2:45and moving forward to 2027

2:48AI actually created way more work for

2:50security researchers.

2:52It might be a bit different work. It

2:54probably evolved.

2:55But still there is a long

2:59way to go.

3:00So why did that happen? In my opinion

3:03there was a wrong perception of AI

3:05capabilities and the sword, meaning

3:07hackers and offense got sharper first.

3:10And the shield, meaning the capabilities

3:12like a silver bullet, magic pill, super

3:15app who closes all your security

3:17questions

3:19is still being built or potentially will

3:21not

3:22be built ever.

3:24So I do think that defenders should pick

3:26up the sword too.

3:28Because again, remember then when you

3:30are trying to

3:32hack the website or the app

3:34you can test one bag across multiple

3:36targets, but when you work as a defender

3:39in a defense team

3:41you have to test for every box in a

3:42single target. That's a bit different

3:45work.

3:46And I don't think that AI made offense

3:50or hackers smarter

3:52or way more powerful. But what I do

3:53think AI did, it made them faster and

3:57exploits way cheaper. So here are the

3:59like four examples.

4:01I won't go like deep into them, but just

4:04to give you a glimpse that, for example,

4:06for ConnectWise,

4:10hackers exploited by the way the

4:12security framework which was published

4:13by security researcher,

4:15and they managed to as quick as 2,900

4:19logins in a minute to try to steal

4:22ConnectWise clients data.

4:25Same similar stuff happened in South

4:26Korea, like in a matter of days, 16

4:28financial institutions, banks and

4:31crypto companies were exploited.

4:33Speaking about the

4:35ease and the like from resource

4:37perspective, the cheapness of hack.

4:39So

4:40Nord Forge VR routers created a great

4:44research and

4:46traced actually one hacker's job who

4:48left a lot of like footprints during his

4:51hacks and turned out that he hacked

4:55attempted to hack 1,200 websites and his

4:57budget was only about $5,000, which

5:00means that it was less than five bucks

5:03per attempt

5:04and

5:05$50 per successful hack actually,

5:09which is

5:10quite impressive. I do believe that

5:12maybe moving forward it will be even

5:13cheaper and we should consider that as

5:15well.

5:16And a bit more speculative stuff, but

5:18Palo Alto Unit 42 says that in PHP

5:22breakout like hackers used AI during

5:27their further trying to stealing data

5:29and exploiting

5:32targets.

5:34Interestingly enough, then I asked our

5:36security fellows

5:38what do they think about the AI usage.

5:43Don't read it as a study. It's not a

5:45formal study. It was just like a

5:47set of conversations

5:48with my buddies and with the network as

5:50well. But surprisingly so,

5:54I had like three almost equal groups of

5:57people

5:58saying that

6:00they have entirely different perspective

6:02or similar but different perspective on

6:04AI capabilities. Like the first group of

6:06people said that actually

6:08we don' think that AI can find anything

6:11that I wouldn't have been able to

6:12identify. Maybe zero to five percent

6:15additional

6:16highs and

6:17criticals. And I'm talking about highs

6:19and criticals, no about low and

6:20informational obviously.

6:22The second group was a bit more moderate

6:24and they said that

6:2510 to 20 percent of highs and criticals

6:28way faster identifiable with AI than

6:31without, but still you have to pay

6:33attention

6:35at least one out of three

6:37highs and creates are missed by AI and

6:38you have to go through the code base,

6:40especially if it's like a

6:42major update to the protocol or

6:45major feature.

6:47The third group was even more optimistic

6:49and they said that they've seen a huge

6:51development of AI capabilities

6:52throughout the year. If in the beginning

6:54it was 10 15 percent additional

6:57value for them, now they see it like 25

6:5935 percent on creates and majors across

7:01the board. It doesn't mean that they

7:03wouldn't be unable to find these bugs,

7:04but it means that they would spend a lot

7:07more time

7:08on it. By the way, I'm really curious to

7:11talk to you after the talk and

7:13see which group is

7:16resembles your view on the AI, so feel

7:18free to approach.

7:20On the first positive side of things, we

7:22have to admit that

7:23AI produces not only the the useful

7:27stuff, but also slope.

7:29And the the sort of on that is quite are

7:31quite interesting. So on first positive

7:33side of things, AI researchers told me

7:35that it's not necessarily bad stuff. If

7:38we get from 30 to 50 percent false

7:40positive rate and we know that we have

7:42to go through this findings manually,

7:44this is a part of the engagement, that's

7:46fine because we get a lot of

7:48interesting leads, interesting security

7:50cues and that will help us investigate

7:53further and understand protocol better.

7:55And maybe this particular issue is not

7:58real, but it will lead me to something

8:00interesting.

8:03On the client side of things regarding

8:05the false positive, Uh the answer was a

8:07bit different. So nice figure from

8:10Coinbase actually. So they reported only

8:134% of their submissions in the first

8:15half of 2026 were real, which is insane.

8:18So 96% of submissions is a complete

8:20fake,

8:21which is a pity, which is a pity. Uh but

8:24we have to understand that there are

8:25different types of noise for clients as

8:27well. So clients told me that if

8:30this report with lots of false positive

8:32stuff goes to the auditing team or to

8:35the security team and we know that

8:37they'll go through it and validate it

8:39and triage it, then it's fine. However,

8:42if

8:42false positive report report with like

8:45lots of false positives goes directed to

8:47dev team to patches and fixes, that's a

8:49real problem. And also uh some clients

8:52told me that they've seen like huge

8:54development in effectiveness of AI and

8:57they received around 50% false positive

8:59rate in the first half of 2026. Now it's

9:02way less, but still much more uh

9:05to analyze.

9:07We've talked about briefly about code

9:09bugs, but actually

9:11uh the majority of money we are losing

9:13right now is not through the code bugs.

9:15Although in quantity they're still

9:17number one.

9:18The majority of uh money lost during

9:22this year were through operational

9:23security, suppliers, partners,

9:26bugs through and not under your like

9:29code control. Let's put it that way.

9:32Couple of examples again.

9:34Like Bitget got exploded because there

9:36was a hole

9:38in the security product they purchased

9:40and they just drained almost $400

9:42million.

9:43Uh another um

9:45one more time for ConnectWise as well,

9:47like a reverse example. They were

9:49holding their clients' data and they got

9:51exposed and

9:532,250 clients got their data leaked.

9:57So it just

9:59matters that

10:01the threat is not only from your code

10:03base. There are multiple venues to got

10:05exposed.

10:07And that brings me to another thought

10:08that you have to treat your security not

10:11as a uh one layer. You have to treat it

10:14as a multiple like layered approach when

10:17in the center you have

10:18your protocol or your uh

10:21client's funds

10:22and you have to build as much layer as

10:24possible around it.

10:26Obviously, it starts off with a design.

10:28Bad architecture, bad design, you'll got

10:30exposed. Then obviously your code should

10:33be uh uh in a good shape. Audits, AI

10:36review, you name it, fuzzing, sometimes

10:39formal verification, everything is super

10:41important here.

10:42But then you have to understand that the

10:45current state of protocols, current

10:47state of market is interconnectivity

10:49between like multiple partners and

10:51vendors. You have to treat all the

10:53integration with your partners almost as

10:55rigorously as your own codebase.

10:57Obviously then goes separation, security

10:59keys and access. Don't forget about

11:01that. On top of that is monitoring. And

11:04uh you have to understand that just

11:06having some monitoring system uh

11:09embedded is not enough. Have you tested?

11:13Did the right person receive uh the

11:14notification when it should? Uh did you

11:18cover the whole potential incidents you

11:20have with your monitoring system? That

11:22the question you have to ask. But on top

11:24of that

11:25you have to have a plan.

11:27Unfortunately as it sound

11:29stuff will break

11:31and in first hours, even days, uh your

11:33brain will not work as efficiently if

11:36something goes sideways. So you need to

11:38have a plan which outlines first steps

11:41you have to do in first minutes, first

11:42hours, first days.

11:44How to stop the drain? How to stop the

11:46leakage? How to notify your partners?

11:48How to make a communication to the

11:50community? Whether you should uh or

11:52should not engage with uh law

11:54enforcement in one capacity. All these

11:56questions should be decided beforehand

11:58before hack happened and every

12:00responsible person should have

12:02this plan

12:03I would say printed out on his desk.

12:07Uh another uh important stuff we as a

12:10security researchers we have to pick up

12:12the

12:13I know it might sound like just use my

12:15AI.

12:16Actually, it's not.

12:18Yes and not.

12:19So, first of all, you have to understand

12:21how attackers are viewing your codebase

12:24and how do they use your

12:26AI to hack you.

12:28And then you have to build upon that.

12:31First of all, you have to build a

12:31process inside the company. How do you

12:34collect the context for your models? How

12:37what standards do you have for teams to

12:39use AI tools?

12:41What happens if you identify issue,

12:42where it goes?

12:44Simple stuff like that, but believe me,

12:46when you have it processed out, then you

12:47can go to the next stage.

12:49What

12:50what is triage for you? How do you do

12:52your triage

12:54in your company? What tools do you use?

12:56Speaking about tools,

12:58how many tools do you use? Do you have

13:00only one in-house tool or you have

13:02multiple tools like third-party tools,

13:04your in-house tools?

13:06How often do you rotate the models you

13:09use? Because we heard multiple times

13:11with our customers that even though one

13:14model could be prominent and one

13:15framework could work well,

13:17the combination always beats it. So, if

13:19you combine this several models, several

13:22tools, several approaches, it will

13:24always

13:26be more efficient than only one

13:28brilliant model you might have.

13:30And remember,

13:32AI alone does not replace an audit or a

13:34security researcher, but

13:37done right, it can show you the way the

13:39attackers will view your codebase

13:41because they will not audit you, they

13:43will try to break you.

13:45And in

13:47the last thought I would like to share

13:49is like security should not be stamp, it

13:51should be a habit, it should be a

13:53routine

13:54you perform every day within your

13:57company.

13:58And obviously,

14:00you don't always have to engage with

14:02third parties and pay a lot of money for

14:04audits and

14:06third-party tools. There are stuff you

14:08can do yourself within your company.

14:10Everyone has access to AI. You can come

14:12up with your own tooling within the AI.

14:14Operational security basics

14:16as simple as sometimes killing your old

14:18accounts as well.

14:20Uh, monitoring, setting up your

14:21monitoring. No one will set up processes

14:23and monitoring for you. Even if you have

14:25like a brilliant report around

14:27operational security from third party,

14:29if you do not engage within your

14:31company, if your leadership is not

14:33interested in implementing this stuff,

14:35nothing will work.

14:37And obviously there is a period of time

14:39when you have to bring the heavy guns,

14:41when you have to engage

14:42with third party companies. I would say

14:45every feature or every change which

14:47involves money moving uh, operations,

14:50new core logic, new features, new bigger

14:52integrations,

14:54or if you want to build like a

14:55operational security playbook for the

14:56first time,

14:58this is the time you have to engage uh,

15:00with third parties because that would be

15:02way more efficient. And

15:05remember like not every change need an

15:07audit, a full audit, but every change

15:09needs someone to look on it.

15:13Uh,

15:14yeah. So getting back to the initial

15:16question, uh, how do you navigate AI

15:18security uh, chaos in 2027?

15:22So the answer is simple. You can't if

15:24you wait uh, for the shield to be built.

15:27But if you lower your defense and pick

15:28up the sword, you have much better

15:30chances uh, to survive and move forward.

15:33That's it for my time. Thank you so

15:35much.

15:36And please

15:37please come to me, talk to me in case

15:39you think I got something wrong. I would

15:41curious to

15:43hear your perspective as well. Thank you

15:44so much.

Source and accuracy details

Source export SHA256: 48f1109120a4055b7d9233b3a75414d6d5ec3e609d188ccce4de2314cf4f3013