Learning Library

← Back to Library

Framework for Securing Generative AI

Key Points

  • Generative AI expands the attack surface, prompting 80 % of executives to doubt its trustworthiness due to cybersecurity, privacy, and accuracy concerns.
  • A security framework is needed that protects every stage of the AI pipeline—data collection, model training/tuning, and inference/usage.
  • Threats to the data stage include poisoning (injecting malicious bias), exfiltration (stealing sensitive training data), and accidental leakage, all of which can degrade model performance or expose confidential information.
  • Mitigation starts with thorough data discovery and classification to identify sensitive assets, followed by controls that safeguard data integrity, prevent unauthorized access, and ensure safe model deployment.

Full Transcript

# Framework for Securing Generative AI **Source:** [https://www.youtube.com/watch?v=pR7FfNWjEe8](https://www.youtube.com/watch?v=pR7FfNWjEe8) **Duration:** 00:13:12 ## Summary - Generative AI expands the attack surface, prompting 80 % of executives to doubt its trustworthiness due to cybersecurity, privacy, and accuracy concerns. - A security framework is needed that protects every stage of the AI pipeline—data collection, model training/tuning, and inference/usage. - Threats to the data stage include poisoning (injecting malicious bias), exfiltration (stealing sensitive training data), and accidental leakage, all of which can degrade model performance or expose confidential information. - Mitigation starts with thorough data discovery and classification to identify sensitive assets, followed by controls that safeguard data integrity, prevent unauthorized access, and ensure safe model deployment. ## Sections - [00:00:00](https://www.youtube.com/watch?v=pR7FfNWjEe8&t=0s) **Framework for Securing Generative AI** - The speaker explains how generative AI widens the attack surface, cites executive trust concerns, and outlines a comprehensive framework that secures data, models, and inference to mitigate privacy, accuracy, and cybersecurity risks. - [00:03:12](https://www.youtube.com/watch?v=pR7FfNWjEe8&t=192s) **Securing AI Training Systems & Models** - The speaker outlines essential safeguards—data classification, encryption, multifactor access controls, continuous monitoring, and strict provenance checks—to protect training pipelines and ensure that imported models are trustworthy. - [00:07:53](https://www.youtube.com/watch?v=pR7FfNWjEe8&t=473s) **LLM Threats: Prompt Injection, DoS, Theft** - The speaker outlines three major attacks on large language models—prompt injection to bypass guardrails, denial‑of‑service overload to cripple performance, and model extraction theft that reconstructs the model from repeated queries. - [00:11:53](https://www.youtube.com/watch?v=pR7FfNWjEe8&t=713s) **Governance and Security for Generative AI** - The speaker outlines the need for a comprehensive governance layer that ensures fairness, bias mitigation, drift control, regulatory compliance, and ethical operation of generative AI, while also leveraging AI to strengthen cybersecurity—creating a reciprocal “AI for security and security for AI” framework. ## Full Transcript
0:00Your attack surface just got a lot bigger.  How? Well, generative AI. This technology 0:06that is taking the world by storm and can do  amazing stuff, also introduces some new threats 0:12with it and new risks. In fact, 4 out of 5  executives have said that they're not sure they 0:19can really trust generative AI because they have  concerns specifically related to cybersecurity. 0:25They have concerns related to privacy and  accuracy of what the system can do. If we 0:30don't address those issues, we don't have trust  and we're not able to take full advantage of the 0:35technology. So what do we need to do in order to  secure generative AI? Well, we need a framework 0:42that will allow us to take a look at these issues.  So let's talk a little bit about what this is and 0:48how it works. Well, what starts is we have data.  We use this data from a lot of different sources, 0:55and we use it to train and tune our models, which  are the next component of this framework. We use 1:03the data to train the model and then ultimately  the model is used for inferencing. The inferencing 1:11is where we get our output. So this is how the  thing generally works. In fact, we have to look 1:16at how we're going to secure each one of those.  How do I secure the data? How do I secure the 1:21model? How do I secure the usage? And it turns  out there's a couple of other things we need to 1:26take a look at that I'll show you before the video  is over. Okay, let's take a look at securing the 1:33data itself. This is the data that we're going  to use to train and tune the model. Therefore, 1:39a bad guy is going to look at that and see what  they can do to it because that's what they do. If 1:44it exists, they're going to try to mess it up.  This is kind of why we can't have nice things, 1:48I'll tell you. And what they're going to do in  this case, one possibility is poison the data.  - 1:55If they poison the data, then basically they  introduce some inaccuracies or things like that 2:01that go then into the training and tuning of the  model. And now the model produces bad results as 2:08a consequence of that. Another thing that could  happen in this space is exfiltration. An exfil 2:15attack would be where someone breaks into this  system and pulls data away from it. And this could 2:21maintain a large set of sensitive information. And  another version of that is leakage. Leakage might 2:29be a more unintentional versus exfiltration is  an intentional where someone has come in. But the 2:35effect is the same. This sensitive data--and we may  use a lot of really sensitive data to train our 2:41model and tune it because that will make it more  accurate and more useful in our use case. But as a 2:47result, now this ends up with a big bullseye on it  that somebody is going to try to hit. Therefore, 2:54we're going to have to try to secure that. How  could we do that? Some things that we should 2:59be taking note of is we should do data discovery  and classification. I need to know where all the 3:06sensitive data is. Maybe these data sources were  secured when they were in their various databases. 3:12But now that we brought all this together as a  training system, maybe we didn't apply the same 3:17controls and we need to. Classify that so that we  know what level of security and protections are 3:24necessary. We want to also use cryptography so  that we make sure that the data that does leak 3:32out or gets exfiltrated out is of no harm. No one  can read it, unless they're an authorized user. We 3:39also add things like access controls, which make  sure that only authorized people can get into this 3:46system using strong multifactor authentication  (MFA) and things like that. And then also we need 3:53to monitor the system. I want to know if this  has occurred so that I can then take remedial 4:00actions and do the right sort of countermeasures  and limit the exposure. Okay, now let's talk 4:06about securing the model itself. Now, how does  this stuff work? It turns out very few people 4:12develop models for their organization because it's  very computationally expensive, requires a lot of 4:18effort. So most people are leveraging existing  models. In many cases, they get them from open 4:24source. And knowing where you're getting your  model from makes all the difference in the world. 4:29So typically I would import models from a lot of  different trusted sources. But if I'm not careful, 4:35maybe I get a model that I think is trusted  and it's not. Maybe it's a copy of a trusted 4:42one that has been modified in some way. And so  that means I really have to consider basically 4:48the models as a supply chain, and I have to do  supply chain management of the models themselves 4:54that go into whatever system I'm going to use. So  in this case, I need to be able to make sure that 4:59I can trust what my sources are, and I need to be  able to know that there are not bad things, 5:06not just bad information, actual malware. There  have been researchers that have done a proof of 5:11concept to show that malware could be introduced  through a machine learning model. So we have to 5:17be able to look for those kinds of things. Also,  in many cases, we're going to use APIs as a way 5:22to communicate out to other services--maybe to  communicate with a model if we're not hosting 5:28it directly ourselves. That API path could also  be a path where an attacker could come in and 5:35introduce some sort of error that then affects  our system. Other things--if we use plug-ins that 5:42elevate privilege, well, I need to be concerned  about the privilege escalation that can occur and 5:48limit what it's able to do so that it's not able  to also make changes here, or put bad information 5:55out, or even modify parts of my system. I want  to make sure that I have a human in the loop to 6:00guard against some of those kinds of things. And  then as well, I've been talking about IT attacks, 6:06but how about an IP attack, intellectual property?  I need to be able to make sure that I'm not using 6:12copyrighted works in one of these sources, because  then I could be sued for using information that's 6:18not really I'm permitted to be using. So what  kinds of things should we be doing to guard 6:23against this? Well, I need to scan my systems  just like I do for other types of malware and 6:29look for any type of of of harmful code that's  being introduced into the system. I need to be 6:35able to harden my system. That basically means I'm  going to remove any unnecessary services or change 6:42all the default user ID and passwords. All of  those kinds of things. The system should be able 6:47to withstand an attack. And I do that hardening by  removing and limiting the attack surface as much 6:54as I can. Then I use things like role-based access  control. Again, trying to make sure that something 7:02like this can't do more than it's supposed to  be able to do. And I can do that on a per user, 7:07per account basis. And then look at my  sources, vet those sources, make sure that they 7:14are trustworthy, make sure that they don't contain  copyrighted materials that will get me in trouble 7:19legally and things like that. So while these are more IT, this is IP. Okay, we're going to take a look at how do we secure the 7:27usage of our generative API? We've looked at the  data, the model--how it's used. And it turns 7:33out one of the main ways that this system can  be manipulated by a bad guy, a malicious 7:39actor, is through what's called prompt injection.  And prompt injection is in fact the #1 7:46on a list of top 10 vulnerabilities mentioned by  OWASP, if you're familiar with that organization. 7:53So prompt injection. In that case, we have a user  that comes in and basically tries to jailbreak the 7:58large language model or generative model that's  behind this by issuing commands, having the system 8:03do things that it was not intended to do. Trying  to get around the guardrails that are put there. 8:09It's almost a semantic attack. It's an attack with  the words on the knowledge itself in many cases. 8:14So we basically with that can also sometimes bias  the output and therefore the information that 8:21comes out of the system is now not as trustworthy  as it should be. Another type of attack someone 8:26may do is a denial of service. If I send enough  attacks into the system that are complex, then 8:34maybe it starts bogging the system down. Because  I can ask a question much faster than a system 8:40can answer it. So we send in a bunch of these and  now the system becomes overwhelmed and it can't 8:45keep up anymore. And then the last one that I'll  mention here is model theft. In this case, someone 8:52might not be able to just break in and steal your  model if you've done a good job up here in the 8:57storing of your data and securing all of that.  But what if they just put a number of different 9:04queries into the system and then take the output  back and do that again and again and again. They 9:10can basically mine the model in order to get  the information that they want. Then they could 9:15potentially go build their own version of this.  And now they've essentially stolen your model in 9:20the process. So what can we do to guard against  some of these threats? Well, one of the things we 9:26should do is monitor. In this case, monitor the  inputs into the system and put guardrails around 9:34those. We'll never be able to put all the kinds  of semantic guardrails that we need to make sure 9:40that someone can never get in. But we need to at  least try and do the best job we can. There also 9:47will probably be a new class of tools that we're  starting to see emerge. One example of this is 9:52machine learning, detection and response. So this  is we've done detection and response and security 9:58for a long time, but something that's specifically  built for machine learning, maybe even generative 10:03models. This is going to be a new emerging area  that we can start looking at. And then finally, 10:08some of our fundamentals using security  information, event management system, or a SOAR, 10:14a security orchestration, automation and response  system that is monitor all of this so that I can 10:20see the abnormal inputs. I can see if the system  is under duress, if it's being overloaded, and 10:27things like that. If someone's carrying data out.  I could use some of those type of capabilities in 10:33order to be aware. So there you have a way that  you can secure the usage as well. 10:39Okay, now for the big reveal of the two elements I told you I  would talk about at the end. AI, generative AI, 10:46does not exist in a vacuum. It runs on IT systems.  Therefore, we have an infrastructure underneath 10:54all of this--traditional computers. And the way  we do security for those is something that I've 11:00talked about before. That's what supports all of  these systems and makes it possible for them to do 11:06what they do. And what I've discussed before, is  this thing called the CIA Triad: confidentiality, 11:12integrity and availability. Those are the concerns  that we have in securing an IT system. Those 11:18concerns don't go away just because now we move  to an AI system. We have to do everything we've 11:23always done, plus a little bit more. And that  little bit more is all of these kinds of things 11:28that I've been talking about in the video. But  we can't ignore the fundamentals of securing the 11:34infrastructure itself and doing the basic blocking  and tackling that we always have to do on every IT 11:40system. And then finally, the last element here is  one of governance. Not so much a security concern, 11:47but it's a big concern for the functional  operation and correct results of the system. 11:53I need to be able to direct, manage, and monitor  how the generative AI works. I need to ensure that 12:01it's fair, that it's not biased, that the model  doesn't have drift over time because someone has 12:06introduced some incorrect information in it.  I need to be able to ensure that that's not 12:11occurring. I need to be able to keep up with  regulatory compliance and requirements issues. 12:16And ultimately make sure that the system operates  ethically. So I need a governance layer here as 12:22well. Put all of these things together and this  is a security framework for generative AI. Now, 12:29if we think about--to summarize--what I've  talked about in another video, is using 12:34AI to create better security. And there are a  lot of possibilities of what generative AI can 12:40do in order to make us better. Make this a force  multiplier for doing the best cybersecurity we've 12:47ever done. And then what has been the subject of  this video is how we can use security to make sure 12:53that the AI is secure. So it's AI for security  and security for AI. If we get all of that right, 13:00then we win. Thanks for watching. If you found  this video interesting and would like to learn 13:05more about cybersecurity, please remember  to hit like and subscribe to this channel.