Your LLM Stack Is a 2008 Database With Better Marketing — Lovina Dmello, NVIDIA

AI Engineer · 20 min · 218 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. 00:01[music]
2. 00:13Hi everyone.
3. 00:14So, my name is Lavina D'Mello and I'm a senior software developer at Nvidia and I work on the deep learning infrastructure team.
4. 00:22So, the title here is a bit of provocation.
5. 00:26What I really mean and what I want you to leave with is uh for the next 15 minutes, I'm going to make one uncomfortable argument, which is almost everything
6. 00:39that is breaking in the production ML security isn't some exotic AI attack.
7. 00:45It's the same boring infrastructure mistakes that we supposedly fixed years ago.
8. 00:52So, now once uh you look across the research from the whole field, the pattern is very easy, like it's very difficult to unsee.
9. 01:02Let me start with a story.
10. 01:04So, into in 2023, security researchers went looking at Ray clusters.
11. 01:12And Ray is one of the most popular framework for distributed ML.
12. 01:17So, what they found out over there was there were thousands of clusters that were sitting open on the internet.
13. 01:26What that means is the dashboards were open and there were job APIs that were open.
14. 01:33And why that happened?
15. 01:35Because authentication was off by default and somebody just forgot to turn it on while putting them ML in them uh in production.
16. 01:53So, I'm a little bit short, you know.
17. 01:55[laughter]
18. 01:57So, I need this.
19. 01:59Thank you.
20. 02:00Okay.
21. 02:01So, uh the exposure at that time was a lot.
22. 02:05Like, it was over billion dollars.
23. 02:08So, now let's sit with this part.
24. 02:10This was not a zero-day attack, and it was not a clever new attack uh on a neural network.
25. 02:18So, what happened was somebody just forgot to turn the default setting on while putting them into the production environment.
26. 02:26So, it just depends upon what the configurations are on different deployment environments in the infrastructure.
27. 02:34And when we look at the breaches that actually make the news, what we see, this is not a rule, it's not an exception, but it's almost never the fancy adversarial
28. 02:44attack.
29. 02:45What is happening is it's maybe like somebody exposed an API key.
30. 02:51What happened is maybe some account was just over-privileged.
31. 02:55Maybe some more model weight was uh in a public bucket.
32. 03:00And the whole breach, it's it was just like somebody forgot to flip.
33. 03:07So, how did we get here?
34. 03:08Let's rewind a little bit.
35. 03:10So, what used to happen is a classic 2008 application was deterministic.
36. 03:16So, we could reason what the code would exactly do.
37. 03:20You know, it had a fixed parameter, and it served one tenant.
38. 03:25Security researchers or security team had a clear mental model.
39. 03:30They knew like what they need to secure.
40. 03:34And like great tools were built around that time.
41. 03:38So, later what happened is we put this machine learning models into production.
42. 03:43And we quietly broke all of the three assumptions.
43. 03:47Because behaviors became probabilistic.
44. 03:50Correct is a distribution now.
45. 03:52It is not one branch like we say yes or no. And then we are protecting the weights which the weights itself are copyable.
46. 04:03And the leaks are straight through the API that we are trying to serve.
47. 04:08And to make it more affordable, we pack everything onto a shared multi-tenant GPU where other security risk come into picture.
48. 04:18And now the LLM stack itself blows up wide open.
49. 04:22And prompts become control flow and answers come from retrieval corpora.
50. 04:28And agents are going to take the actions for us.
51. 04:32So here is the thing.
52. 04:34We changed everything about the stack when we moved to machine learning, but we forgot to change the security assumptions.
53. 04:42So we are running a 2028 playbook on a 2026 system.
54. 04:47That's what the whole talk is about.
55. 04:50So few things, five things stand out across all of these and I have tried to boil them down into three core ideas.
56. 04:57Uh and here they are.
57. 05:00So first one is misconfigurations.
58. 05:02So lot of the times misconfigurations happen.
59. 05:05It is not the missing features that are number one failure reason.
60. 05:10Second is the defenses that we are actually reading, they might not actually survive into production.
61. 05:17And third one is we have to secure ML like infrastructure and it's not like securing just a model, you know?
62. 05:25And if we if you forget everything about this talk, I want you to remember these three points over here.
63. 05:32Uh and I'll share a few couple of things at the end.
64. 05:35Uh Okay.
65. 05:38So over here is a map I want you to carry with you.
66. 05:41It's four pillars about the defenses in depth.
67. 05:46So at the bottom we can see the at the bottom in the black we can see the infrastructure security.
68. 05:52Over here comes into picture the containers, the networks, the GPU virtualization and encryption.
69. 05:59On top So, this infrastructure security is the foundation.
70. 06:04And on top of it sits the access control like who is allowed to do what and then is the runtime security which is checking uh and watching the traffic while the system is life.
71. 06:19So, they feed each other and wrapping all of it uh to day-to-day practice like um how the team works plus compliance, incident response, and honestly the culture.
72. 06:33So, one thing to remember over here is if uh infrastructure is the foundation and failures down there can cascade up.
73. 06:42So, if we have a perfect access control policy, it's just a decoration if the cluster underneath is wide open.
74. 06:51So, yeah.
75. 06:53So, those are the four levels.
76. 06:55And this is a quick map of what we are defending against.
77. 07:00So, there are six categories.
78. 07:02The two are about the model itself.
79. 07:05So, where we have the adversarial inputs and we have the model extraction.
80. 07:10Over here the primary defenses should be uh about how we do the input validation, how we do the adversarial detection, how we do the rate limiting, API auth, and query analysis.
81. 07:23The next two are about the data and integrity.
82. 07:26Over there comes the poisoning and supply chain and privacy breaches.
83. 07:31So, we should make sure that over there we make sure to secure the CI/CD, model signing, and provenance, access control, encryption, all of those comes into that picture.
84. 07:43And the last one, our classic infrastructure, where we take care of um basically container isolation, network segmentation, uh R back, and behavior analysis.
85. 07:57So, if we notice one thing over here, uh the green ones, infrastructure compromise and insiders, those are exactly where the real breaches keep landing.
86. 08:09And they are the cheapest to get wrong by accident, which brings me to the core idea number one, which is mis- misconfiguration, not missing feature.
87. 08:18So, those that is one of the number one failure mode.
88. 08:22Um okay.
89. 08:24So, let me walk you through one of the research.
90. 08:28What happens is researchers audited 50 real production setups running machine learning.
91. 08:35In 78% of them, what the researchers found out was at least one critical security mistake.
92. 08:43And and And here's why that mattered beyond engineering.
93. 08:48So, that was a misconfiguration, and the difference between a normal Tuesday and something having a headline about our company in the news is this kind of incident, where there could be like a critical misconfiguration
94. 09:03mistake.
95. 09:05And And the same three problems kept showing up in this research study.
96. 09:11First was access controls were left wide open.
97. 09:15So, what happens when an access controls are left wide open?
98. 09:19Almost any account can do almost anything.
99. 09:23So, that is one of the problem.
100. 09:25Second is nothing sep- nothing was separated.
101. 09:29For example, different parts of the system So, once an attacker would get access to one part of the system, he could do anything in the whole system.
102. 09:39So, there is the thing like separating different parts of the system.
103. 09:43And the third one were the passwords and trained models.
104. 09:49So, themselves they were sitting into a storage that anybody could reach.
105. 09:54So, why does this keep happening?
106. 09:55So, because this production protections are built by security experts for security experts and then they get handed to the ML teams.
107. 10:05And the responsibility of ML team is um basically making the model accurate, not configuring the infrastructure.
108. 10:14So, the protection is right there in the platform.
109. 10:17It just sometimes get switched off by some configurations.
110. 10:25Okay.
111. 10:26So, now moving to the core idea two, which is defenses you read about don't survive into the production.
112. 10:33So, here is here is the thing.
113. 10:37Every control costs us something.
114. 10:40And in ML that something is latency and throughput.
115. 10:45The two things that SLA is made up of.
116. 10:49Let's watch the colors.
117. 10:51So, the green tier over here, the basics like logins, input checking, costs do those cost under about 8% and that should always be there.
118. 11:03No debate over there.
119. 11:04The second one is the orange tier.
120. 11:07So, this is like the heavier isolation that walls around that that walls workloads of each other.
121. 11:16So, here it runs 10 to 20%.
122. 11:19So, we can use it selectively like for sensitive and untrusted workloads, what we can do is we can decide if we want it or not.
123. 11:28And the purple ones is catching malicious input in real time.
124. 11:31So, this is the most expensive one.
125. 11:34Over here, it can cost like I don't know, 15 to 30% and it's one of the thing that researchers love the most.
126. 11:42But, because we cannot slow every request down, so it's a non-starter.
127. 11:49But, what really matters is how much it cost depends almost entirely on how well we build it.
128. 11:56So, the same security check check, it can be done sloppily and it can double our response time.
129. 12:04But, if it is done well, like considering all of these things, then it's fair, like the balance is fair.
130. 12:11So, the guidance team actually need isn't should I do this, it's like how do I do it without slowing everything down.
131. 12:20And then we can run the cheap controls everywhere and we can save the expensive ones for the heavier risk systems.
132. 12:31So, here is the the next slide.
133. 12:34The ecosystem isn't production ready yet.
134. 12:36So, what I meant is um So, there is the first gap that we will see.
135. 12:43The security tooling for ML is years behind the rest of the software.
136. 12:49So, regular software teams, they solved a lot of this decades ago with automated security scanning, proper passport password management, and then we haven't carried it all over to ML yet.
137. 13:04Even today, there's a good share of early warning systems for ML specific threats.
138. 13:10So, the second gap is in the expertise silos.
139. 13:13So, security team don't speak ML, ML teams don't speak security, and the operation teams don't know how the model behaves.
140. 13:22So, there is a gap over here and the problem leaves in the gaps.
141. 13:26So, between those gaps where nobody has a full picture.
142. 13:30So, there's a little bit of mismatch over here.
143. 13:33And underneath both is a gap between the research and reality.
144. 13:37So, research is busy defending against tiny invisible tweets.
145. 13:44But to an image that while the real world, we are fighting against the stolen password and then people are misusing access that they already have and it studies one model on its own while we are running dozens of them
146. 14:01on shared system.
147. 14:03So, what it lacks are one team can actually deploy everything.
148. 14:09So, there is this core idea number three.
149. 14:11It's the constructive one.
150. 14:13So, if the problem is infrastructure, we have to secure it like infrastructure, not like a model.
151. 14:18And two practical tools for that are a maturity model to know where we stand and a short list of how to fix it first.
152. 14:31So, here is the next slide.
153. 14:34What we have to do is how to figure out where we actually stand.
154. 14:38It's a maturity model that will help us map onto the NIST AI risk management framework.
155. 14:45So, there will be four levels, each tied to an overhead budget and the context.
156. 14:51So, on the level one, we can see that under 5% overhead is the bare basics.
157. 14:57Fine for test environment, not for production.
158. 15:00So, the level two 5 to 10% is a real baseline like proper logins, encryptions, separating our networks, basic monitoring.
159. 15:10This is the minimum bar for production.
160. 15:12And the level three adds more advanced controls that regulated industries like health care and finance needs.
161. 15:21So, on this part, most of teams teams believe that they are at level three, but uh they are actually either at level one or two, like it really depends
162. 15:31how we have set up everything.
163. 15:38So, on this slide, I want to discuss about the three misconfiguration behind most of the risk.
164. 15:44So, the default setup often gives every account a full admin rights.
165. 15:50So, that is what have been observed in most of the research studies.
166. 15:53With credentials that never expire, so there are there have been credentials that do not have any expiry.
167. 16:00And that can leak the key on its own.
168. 16:04And like each account like should do only what it needs to do.
169. 16:11And then the credentials to should expire very quickly.
170. 16:17Now, nothing stops one part of the system from reaching another part of the system.
171. 16:21So, if that can be tackled in a way uh then that can be solved.
172. 16:27Like attacker gets in one part of the system and then they get it can do everything.
173. 16:32So, that's uh like a flat network kind of issue.
174. 16:37Then the fix.
175. 16:38So, what we can do to fix this is we can wall the parts from each other.
176. 16:44And we can require a verified identity between services, not just network access.
177. 16:50So, there is an another issue called the exposed secrets and models.
178. 16:55So, over there passwords could be hardcoded in the code, model files sitting in the open storage.
179. 17:02So, the fix could be like a proper secret manager, encryption, and automatic scanning before before all of this ships.
180. 17:10So, this is nothing but like a basic hygiene.
181. 17:13And this kind of hygiene can prevent breaches that actually make the news.
182. 17:21So, uh here's the practical stuff.
183. 17:24Uh so, now a couple of personal takes on where this is heading because the ground is moving very faster under us.
184. 17:31So, this is the part uh that's least settled.
185. 17:34And four things are on my mind right now.
186. 17:37So, one is prompt injection.
187. 17:39So, the model can't reliably tell the difference between our instructions and someone else's input.
188. 17:49So, a cleverly worded message can quickly hijack the whole system.
189. 17:54So, rag poisoning, if our model pulls answers from a document store, and if the document store is poisoned by somebody or somehow, then what happens is it can poison like everything.
190. 18:09And nobody has to retain anything, right?
191. 18:11So, GPUs So, next is the GPU side channel.
192. 18:15So, where we can pack multiple customers on one GPU in order to save money.
193. 18:20That time information can leak from one customer to the other.
194. 18:25And another is the supply chain.
195. 18:27So, what happens over there is we download models and add-ons from public sources without uh really checking where they come from.
196. 18:37So, that's another issue.
197. 18:39And the honest caveat over here is uh all of this moves very fast.
198. 18:45Every week something or the else comes up.
199. 18:48So, we treat it as a direction, not a gospel, and the defenses are immature right now because uh the threats keep evolving.
200. 18:59Okay.
201. 19:00So, three uh So, let's bring it together.
202. 19:03If you forget everything I said today, here are the three takeaways.
203. 19:07So, one is securing ML is an infrastructural problem, not a model problem.
204. 19:13So, we have to spend our budget where the breaches actually are happening.
205. 19:19Second, the field has enough defenses.
206. 19:22It just need deployable ones.
207. 19:25The frontier isn't a new attack defense pair.
208. 19:28It's making the existing controls run with proper production overhead.
209. 19:34And that's the engineering problem, which is a good news for a room full of engineers.
210. 19:38And the third one is we have to fix the misconfigurations, like over-privileged accesses, flat networks, and exposed secrets.
211. 19:47So, then climb the maturity model on purpose.
212. 19:52And to close the loop on the title, your LLM stack really is a 2008 database with better marketing.
213. 20:00So, secure it like a database, lock down accesses, segment the network, and protect the data at rest.
214. 20:09So, we already know how to do that.
215. 20:11We just have to do it actually in a better way.
216. 20:14That's all.
217. 20:15So, thank you.
218. 20:18[applause]