AI Hackers Are Faster Than Your Pen Test — Eli Cohen, Snyk

AI Engineer · 19 min · 225 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

1. 00:12Great.
2. 00:12Hi everyone.
3. 00:14We're going to talk today about continuous offensive security.
4. 00:17Okay.
5. 00:17I'm going to speak quite soon.
6. 00:19What does it mean?
7. 00:20Uh but before that, a little bit about me.
8. 00:22I'm very excited being here today.
9. 00:24I'm Ellie.
10. 00:25Uh used to be co-founder and CEO of Helios, a runtime security company.
11. 00:29got acquired by sneak and in the last two and a half years I had the privilege in working in sneak primarily on product management for AI products but more recently
12. 00:39founding field CTO meaning I'm working with developers and with security folks to better understand and to help them shape their strategy around AI security.
13. 00:49So how do you take your AI agents and making sure they are um secured?
14. 00:55How do you work with your agents in production?
15. 00:58Um, back there, can you set up the timer for me, please?
16. 01:03The timer.
17. 01:04So, I'll know the time.
18. 01:06Great.
19. 01:07Okay, now we have the timer also.
20. 01:08Great.
21. 01:09Um, so just so I'll get to know a bit better.
22. 01:12Please raise your hand.
23. 01:13Who is here more from the development side of the table?
24. 01:17Developers.
25. 01:19Okay, great.
26. 01:21From the security side of the table.
27. 01:22Yeah, I already know you.
28. 01:24Great.
29. 01:25and other people like sales, marketing, other people running agents in production that are not from developmental security.
30. 01:31Great.
31. 01:32Okay.
32. 01:32So, we have everyone here.
33. 01:34So, I'm going to tailor this uh session towards everyone.
34. 01:37So, obviously this is an AI engineering conference, right?
35. 01:41So, we already know that the amount of code generated today is incredibly bigger than it used to be before.
36. 01:49Okay.
37. 01:49According to our data, it's 218% more lines of code per the developer than it used to be before.
38. 01:58Okay.
39. 01:59And more than 85% of the developers are actually using coding agent now probably running here today.
40. 02:06It's more than 100% right because I guess that each one of you probably have five different session of cloud code running simultaneously.
41. 02:13Otherwise, you wouldn't be here today.
42. 02:16And the funny thing about those agents running in production and generating fork for us is that it's far less secured than we are used to.
43. 02:24Okay, so 62% of the output generated by LLMs is not secure or it's broken.
44. 02:32Okay, so just imagine this data combined together.
45. 02:34It's meaning that we are generating more code than we ever did and this insecure code is getting into production.
46. 02:43So that's a huge challenge for us as a community and what it actually means is that our backlog of security issues is becoming longer and bigger and much more challenging
47. 02:56to remediate it.
48. 02:57Okay.
49. 02:57So we are actually creating many more issues at a pace that we simply couldn't remediate or mitigate.
50. 03:05And not only that, AI is helping our attackers also.
51. 03:10So they are able to chain low vulnerabilities together and to form critical vulnerabilities from those um low low severity vulnerabilities.
52. 03:21And what's actually is going on is that our attackers are leveraging those LLMs and they're attacking us especially where it's very difficult using um application context.
53. 03:33This is actually where our LLMs are struggled the most because they are creating many more vulnerabilities around those areas.
54. 03:41So obviously we are leveraging AI but our attackers are also leveraging frontier models.
55. 03:48Okay.
56. 03:48And the numbers are simply incredible.
57. 03:50Okay.
58. 03:50We are familiar with an operating model that actually was able to identify more than 600 vulnerabilities uh to breach 600 firewalls in more than 555 countries.
59. 04:03Okay, this is simply unbelievable.
60. 04:06And the time for an AI attack to be effective is constantly getting faster and faster.
61. 04:13Okay, so the average time for an AI attack to be successful is 24 34 minutes.
62. 04:21Okay, well the fastest time is 4 minutes.
63. 04:25Okay, it means that since the moment I began talking, there is probably an AI attack out there that was able to successfully breach a production system and 43% of all the MCP servers
64. 04:38actually has vulnerabilities.
65. 04:40Okay, so combining all of this data together, what we realize is that we are producing much more code than ever.
66. 04:48Each one of us has a few sessions simultaneously of cloud code of codex.
67. 04:53Those this code is insecure.
68. 04:56It's getting into production and the people we are facing the attackers are actually leveraging the same tools and they are acting faster than ever in trying to actually leverage those vulnerabilities
69. 05:09and our backlog is simply getting piled up and we are not able to get it.
70. 05:13And this is exactly what I want to talk about today.
71. 05:16what we should be doing to to help improve our security posture.
72. 05:21And this is ties heavily to how we can leverage offensive security not just for the attackers but for us as defenders.
73. 05:33So why traditional security can't really keep up?
74. 05:36So there is a mixed audience today both people from the development side also people from the security side.
75. 05:42So I want to cover first a few basic concepts okay to make sure everyone is familiar with.
76. 05:47So what used to be really easy about operating was static code scanning okay SAS.
77. 05:53So imagine that you are pushing new code to production.
78. 05:56There is a static scanner that is scanning that and is looking for SQL injection cross-ite scripting.
79. 06:02So it used to be relatively cheap and relatively easy to run this across every code change.
80. 06:08And this is great.
81. 06:09Now what it doesn't catch, it doesn't catch things that are going into runtime.
82. 06:13Okay, so if you have any authorization issues, if you have any configuration issues, for that you will need dynamic application security testing.
83. 06:23Okay, you'll need something that test application dynamically.
84. 06:26One example is bola uh broken object level authorization which basically it means that user A can access the data of user B. Okay, those kind of vulnerabilities you can never find
85. 06:40with static analysis.
86. 06:42For that you need to prop the application in runtime.
87. 06:45So what actually a dust scanner is doing it just takes a predefined set of tons of payloads and it just sending to the application to the different API endpoints
88. 06:56and by that it managed to find vulnerabilities.
89. 07:00So this was also something that could find new things but it wasn't good with the biggest challenge we have with the code that LM are producing and this is the business logic.
90. 07:11So for that we've been you what we've been using as an industry it's the pentesting.
91. 07:16Okay.
92. 07:16So what's the idea about pentesting?
93. 07:19It's a human pentester that really understands the concept and the business logic of the application and is trying to break it and to find all the vulnerabilities.
94. 07:30Now it is the golden standard in application security testing but it has some challenges right.
95. 07:37First of all, it's cost a lot of money, right?
96. 07:39Because you actually needed to have a security researcher, a human, not a human in the loop, a human that is actually doing the testing and trying to break the application.
97. 07:49And that's cost a lot of money.
98. 07:51Now, the second problem is because it cost a lot of money and because the process was so slow, companies can only do it once a year, two times a year,
99. 08:02maybe three times a year.
100. 08:04Okay?
101. 08:05So combining the fact that nowadays our AI attackers are trying to continuously attack us like doing this pan testing two times a year is simply not good enough.
102. 08:16Okay.
103. 08:16So what you going to do with the rest of the 350 days you have a year that no one tested your application.
104. 08:22Not now not only that the results you got the result you got this crazy and glorified report in PDF that you had to send to your developers.
105. 08:32there to take it back to understand what's real and what's not and then to start fixing that and then you had to retest.
106. 08:39Okay, it sounds like a movie from the 90s, right?
107. 08:41It doesn't make sense to do this anymore.
108. 08:44But this was the Gund standard.
109. 08:45This is what we were doing.
110. 08:47Now the thing is this is the best method out there to actually test your application for its business logic.
111. 08:56Okay.
112. 08:56So what changes today?
113. 08:58Today we have a new technology and this is exactly what I want to talk how we can leverage this new area of AI tools and LLM and frontier models to actually do pentesting
114. 09:11on a continuous manner in a scalable manner and this is why we're introducing EVO continuous offensive security.
115. 09:21Okay, so this is obviously very tied up to what we're doing within SNIK, but what I'm going to tell you today is general.
116. 09:29Okay, you can take the principles of what I'm telling you and you can actually verify that with any vendor out there in the expo.
117. 09:37Okay, so I want to I want you to think about the following things.
118. 09:41Okay, so what you need to do today, you need to prepare your production system from the unknowns unknowns.
119. 09:48Okay, and for that you need a new paradigm.
120. 09:50Okay.
121. 09:50So first of all obviously you need the AI pentesting capabilities which I will double click on soon again.
122. 09:56But you also need to start taking the concept of red teaming.
123. 09:59Okay.
124. 10:00So agent red teaming the idea is to take multi-steps attacks.
125. 10:04Okay.
126. 10:05So to try to simulate prompt ejection or data excfiltration and goal aenting and obviously to combine it with dynamic testing that is good at identifying authorization issues.
127. 10:15So we are taking all of those capabilities together.
128. 10:18We are building one suite out of that for offensive security as part of our EVO offering.
129. 10:24And the idea is to help you actually mitigate or help you uh handle the fact that the attackers are leveraging the same tools that we are using and they are doing that in a continuous manner
130. 10:3824/7 finding vulnerabilities faster than before and your backlog simply cannot can stand up with that.
131. 10:45So what we're doing differently within AI pentesting.
132. 10:49Okay.
133. 10:50So we I told you before the great promise of pentesting was always the fact that the human and the end really understand the business context.
134. 11:00Okay.
135. 11:00So now we are taking this to the LLM and if there is something that LLM are very good at is really trying to understand the business context.
136. 11:07So behind the scene we are leveraging the LLMs and we are doing that in a continuous manner.
137. 11:12Right?
138. 11:12So no longer you're going to do things like do pent testing twice a year or three times a year.
139. 11:19We're going to do it on every code change.
140. 11:21Okay?
141. 11:22Every PR that you send, every delta, we're going to run this pentest for you and we're going to tell your developers what do they need to fix and how do they need to do it differently.
142. 11:31Obviously we leverage our understanding of the application context, but we also help reduce the false positive.
143. 11:38For us as a industry, we have a severe problem of false positive.
144. 11:42So we are actually proving you that the vulnerabilities that is that we are finding are actually exploitable and we just just flag them as bugs.
145. 11:52We actually help you prove that they are exploitable and we take contexts for many different machines and for many different environments and we combine them together to help guide our LLMs.
146. 12:05Okay, so this is what we're doing differently and I think this is really what's makes this solution so unique and so well positioned to the fact that each one of us have a few coding gadgets running in production and pushing more code than ever.
147. 12:20Now how do we do it?
148. 12:21How does it work behind the scenes?
149. 12:23Okay, so what I want you to understand is we are not taking the static scanners or the dust scanners and just combine him with LLM.
150. 12:32Okay, we build it from the ground up using LLM.
151. 12:36So we have a series of a few agents working together.
152. 12:40The first one is the LLM orchestrator.
153. 12:42Okay, imagine it as the brain.
154. 12:45This agent is working according to a dynamic plan that it follows and it knows how to operate all the different sub agents.
155. 12:53The first agent or actually the second agent is the recon.
156. 12:57Okay, you can imagine it as your elite intelligence unit.
157. 13:01The idea is to collect all the data it can about the different network endpoints, the different APIs, the different fingerprints, collect all of this data together, formalize an architecture out of that and hand it to the next agent.
158. 13:17And the next agent is the vulnerability testing.
159. 13:20It's actually again a series of a few specialized agents.
160. 13:24Some of them are targeting the well-known classes of vulnerabilities.
161. 13:28Some of them are actually hunting for vulnerabilities in the wild and they are trying to break your application to understand the business context and to understand where can attacker can get from
162. 13:39and then they hand it into the judge as I like to call it the exploit validation.
163. 13:44So actually every vulnerability that the previous agent find are now being handed over to this judge and the goal of the judge is to tell us whether this is a real vulnerability or or not to tell us whether this can be exploitable or not.
164. 14:01Exactly because we want to reduce the cognitive load.
165. 14:04Exactly because we want to reduce the amount of false positive we have in the system.
166. 14:10Coming up next is the remediation agent because everything we do we want to make actionable.
167. 14:15Okay, we don't want to report about a vulnerability that eventually the developers they don't know what they can do with that.
168. 14:22And coming after that is the reporting agent because everything needs to be reported.
169. 14:31So what makes us different and those are the point I urge you to check with every vendor out there trying to provide you with the solution or whether you want to implement a solution inhouse for AI pentesting.
170. 14:43This is a huge industry right now.
171. 14:47So I think the most important thing you need to remember is and you already know that especially the developers in the room context is everything.
172. 14:56your LLM are as good as the context that you feed them with.
173. 15:00So what we're trying to do, we are not just taking our AI pen testing solution as a standalone.
174. 15:07We actually feed it with data from all the previous scanners.
175. 15:11Okay, so the the statics code scanner, the dynamic code scanner, all of the different agent engines are being combined together and they are being fed into a our AI pentesting solution
176. 15:24because that way we know how to tailor and we know how to guide the LLM where to look for and as you know also for mythos or for frame
177. 15:33when those models know where to look the results are significantly better.
178. 15:39So that's one thing we're doing that is you know very unique and very differentiated.
179. 15:45Um and also what I want you to remember is that one solution is not enough.
180. 15:51Okay.
181. 15:51The great thing about it is when you combine those things together and based on the data that we see both from the labs and from real data from customers
182. 16:01combining the data together is really what's important and you can't really you can't really use just one of them.
183. 16:08Okay.
184. 16:08There are certain classes of vulnerabilities.
185. 16:12This dust is the best scanner for that.
186. 16:14Okay?
187. 16:15Because you have a set of really exhaustive pre um a set of really exhaustive uh payloads that you can send to your application.
188. 16:24So it's very good for example with quite scripting.
189. 16:27Okay.
190. 16:27So dust for crossite scripting is the best solution.
191. 16:31But if you want again like business logic then you should combine it with uh pentesting because it will be the best solution.
192. 16:39So I think what I really want you to take from this, if there is one thing I want you to take from here is to realize that the world is changing.
193. 16:49You already know that.
194. 16:50That's what made you come in.
195. 16:52But you know we used to test our application once every few months because the change and the velocity of the change was so slow and we're we had people at the end that were reasoning about the application
196. 17:07and about the change.
197. 17:09But that world is does not longer exist.
198. 17:11Right?
199. 17:12Right now we are pushing code to production 24/7 and we have LLMs that are trying to reason about the application context and they are trying to break it continuously.
200. 17:23Some of them can do it in like as fast as four minutes.
201. 17:26Okay, that's almost five time of the lecture of the session that I just did.
202. 17:30So when you're coming to look for new solutions for this problem, you have to ask yourself four questions.
203. 17:36Okay.
204. 17:37First of all, do we run this solution in a continuous way or is it just a point solution that we run every once in a while?
205. 17:46Okay, that's a major difference.
206. 17:48Second, do they really reason about the application context?
207. 17:52Okay, because given the nature of those LLM, application context is really where the new vulnerabilities are coming from.
208. 18:01And third, what more context can they digest?
209. 18:06Okay, how can they enrich the data?
210. 18:08Because the richer the data is, the better and by the way cheaper the LLM operation is going to be.
211. 18:15And the fourth part is do they really help me prove that there is a bug and to make sure that this is exploitable or they they just flag it as a bug.
212. 18:27Okay, so our goal is not only to help you find the bugs faster, but to actually do it better and cheaper and to help you as a developer or as the security people here
213. 18:39focus on what matters most.
214. 18:42Um, so thank you very much and uh I used to do it in every lecture.
215. 18:45You know, my my my dream was to be a rockstar, but I had to be a cyber entrepreneur.
216. 18:50So instead of that, I'm going to take pictures.
217. 18:52So hands up everyone.
218. 18:56Okay, great.
219. 18:58Thank you so much.
220. 19:01And I'm here for questions.
221. 19:02If you want, go for the boot.
222. 19:04It's very close here.
223. 19:05sneak.io.
224. 19:06Meet Evo, our new AI security offering.
225. 19:09Thank you so much.