Learning Library

← Back to Library

Red Team Tackles AI Threats

Key Points

  • AI introduces an entirely new attack surface, requiring security teams to continuously learn and adapt to novel threats rather than treating it as a one‑time testing effort.
  • Chris Thompson leads IBM X‑Force’s Red Team, which comprises about 180 hackers who focus on advanced penetration testing for high‑value targets such as banks, defense contractors, and nuclear facilities, and they actively share tools and research with the wider security community.
  • Rapid AI adoption across products is outpacing proper risk assessments, leading many organizations to deploy generative AI and machine‑learning systems with insecure configurations, such as missing authentication and unsafe code execution in production.
  • These misconfigurations allow attackers to access sensitive assets like enterprise data lakes, execute arbitrary code, and bypass traditional security controls, amplifying the potential impact of AI‑related breaches.
  • Ongoing vigilance, thorough security testing, and a mindset that treats AI systems as continuously evolving targets are essential to mitigate the heightened risks introduced by generative AI.

Full Transcript

# Red Team Tackles AI Threats **Source:** [https://www.youtube.com/watch?v=HZYBj-zeUlY](https://www.youtube.com/watch?v=HZYBj-zeUlY) **Duration:** 00:22:27 ## Summary - AI introduces an entirely new attack surface, requiring security teams to continuously learn and adapt to novel threats rather than treating it as a one‑time testing effort. - Chris Thompson leads IBM X‑Force’s Red Team, which comprises about 180 hackers who focus on advanced penetration testing for high‑value targets such as banks, defense contractors, and nuclear facilities, and they actively share tools and research with the wider security community. - Rapid AI adoption across products is outpacing proper risk assessments, leading many organizations to deploy generative AI and machine‑learning systems with insecure configurations, such as missing authentication and unsafe code execution in production. - These misconfigurations allow attackers to access sensitive assets like enterprise data lakes, execute arbitrary code, and bypass traditional security controls, amplifying the potential impact of AI‑related breaches. - Ongoing vigilance, thorough security testing, and a mindset that treats AI systems as continuously evolving targets are essential to mitigate the heightened risks introduced by generative AI. ## Sections - [00:00:00](https://www.youtube.com/watch?v=HZYBj-zeUlY&t=0s) **Inside the AI Security Mindset** - The conversation examines how AI creates a novel attack surface that demands continuous learning and testing, with insights from IBM X‑Force’s red‑team leader and an AWS senior security architect. ## Full Transcript
0:00being in security you're constantly 0:01having to learn new technology so that 0:03you can break it so you can understand 0:05how it works under the Ood and AI is is 0:08no exception the the biggest piece for 0:10us is it it's not just you know a new 0:12web application framework or you know a 0:15new web server technology it's a brand 0:18new attack surface with a lot of really 0:20uh interesting new new attacks that can 0:23be conducted against it and so it's not 0:26something that you test months and 0:27you're 0:28done no nobody's interested in security 0:31until something breaks and then it's all 0:34we can think about so on today's episode 0:36I want to get inside the AI security 0:38mindset with two people who know it best 0:41one is a real life hacker who just 0:43happens to lead a security team at IBM 0:46called xforce yeah you heard what I said 0:48xforce the other a senior security 0:51architect at AWS joining me now Chris 0:54Thompson what's up Chris hey thanks for 0:57having me Chris welcome to the show I 0:59got to ask you how did you get here I've 1:02had the great privilege and honor to 1:05build the uh adversary services or the 1:08red team at xforce Red over the last 1:10eight years um so that team is really 1:13focused on hacking into big Banks or 1:16defense contractors or nuclear power 1:19plants and and you know showing how we 1:21can achieve objectives and not get 1:23detect it can you tell me a little bit 1:25about about the people that you work 1:27with there at X Force raid we have over 1:29I think 180 hackers here everyone 1:32contributes whether that's in 1:34vulnerability management or application 1:36security testing we're on the red team 1:39last year for example we spoke at 1:41blackout security conference six times 1:43um we put out some really Innovative 1:45post exploitation tooling um a lot of 1:48open-source security testing toolkits 1:50for the community to use Chris we've 1:53we've actually seen some things get 1:54broken recently you know a lot of big 1:57security breaches how has Genai changed 2:00hackers are doing and how you work being 2:02in security you're constantly having to 2:04learn new technology so that you can 2:06break it so you can understand how it 2:07works under the hood and AI is is no 2:10exception the the biggest piece for us 2:12is it it's not just you know a new web 2:14application framework or you know a new 2:18web server technology it's a brand new 2:20attack surface with a lot of really uh 2:23interesting new new attacks that can be 2:26conducted against it and so it's not 2:28something that you test and you're done 2:31especially the risk is compounded 2:33because of how many companies are 2:35rushing to slap AI into every product 2:37Under the Sun and they're trying to you 2:39know have that first mve mover advantage 2:41and they're not pausing to really 2:44perform proper risk assessments so we're 2:46finding when we test these these gen 2:49applications and these traditional ml 2:51you know use cases is we're we're seeing 2:54a lot of these environments have been 2:55set up without proper 2:57authentication um they're allowing codec 2:59execution in the backend um production 3:02environment they're allowing us to get 3:04access to Enterprise data lakes and you 3:07know it's really just a symptom of 3:08people trying to move way too fast um 3:11and worry about security problems later 3:13what does a day look like at X Force 3:16typically you know we we would hear from 3:18a customer that it's about to roll out 3:20you know a new financial application or 3:23new foreign exchange platform or maybe a 3:25new banking platform and they'd say hey 3:27we're we're integrating this chatbot so 3:29a customer can prer a wire transfer or 3:32look up account history for example and 3:35so we're looking at from a traditional 3:38web application perspective what are 3:40these traditional attack paths that we 3:42need to be concerned of and what are the 3:44new uh types of attacks that we need to 3:46worry about when these multimodal chat 3:49Bots and these multi-turn chat Bots that 3:51that are interfacing with different apis 3:53in the back end so we we need to figure 3:55out hey how how is this functionality 3:58being called couldn't I talk 4:00really ask this chatbot to load um 4:03insecure code from the internet could 4:06they build you know some sort of code 4:08execution flow with within the B chatbot 4:11itself um and so there's a lot of of new 4:14attacks that we really need to be aware 4:16of and threat model as we're testing 4:18these applications to ensure that you 4:21know we're not seeing customers end up 4:23in the news when these companies are 4:26rushing to integrate AI you mentioned 4:28just how dangerous it can be if you move 4:31too quickly with it what percentage 4:33would you think are actually secure 4:35that's really tough of I I think there's 4:38there's so many components to leveraging 4:42AI into a solution you have the models 4:47that go into these applications that are 4:49being trained and tuned and you know 4:51downloaded from hugging face and 4:53evaluate it you have the model training 4:56and tuning environment itself you have 4:58the whole m SE Ops pipeline which is new 5:02what's your first step when you're 5:03starting to hack a new gen project it's 5:05taking a kind of a risk-based approach 5:07to threat modeling so is this just a a 5:10you know a sales chap bot that's going 5:12to help somebody you know pick a widget 5:14off of a of a a storefront or is this 5:18you know a healthc care application 5:20that's going to allow you to query 5:22previous um diagnosis from your medical 5:25history right and so we start to think 5:28about how how sensitive is the data 5:31that's being pulled into this 5:33application from Enterprise production 5:35environments and then what um you know 5:38different systems is this interfacing 5:40with is this making calls to like an 5:42ethic Healthcare database is this making 5:45calls to like a customer relational 5:47management system how are those apis 5:50secured you know what's being exposed 5:52publicly what's privately what can we 5:54hit from within the Gen solution and 5:57then we come up with a testing plan we 5:59say okay you know this this type of 6:02testing is appropriate for this whereas 6:05you know this this we can spend a little 6:07less time on we also try to push a lot 6:11of teams to not just think about the 6:13front-end solution because we're seeing 6:16a lot of of code execution and platform 6:20compromise and and you know a lot of 6:22it's private and not hit the news and we 6:24just happen to hear about it from 6:26different customers or from Partners we 6:28think it's really important that that 6:31customers start to think about you know 6:34again that endtoend flow of how did we 6:37evaluate these models how do we know 6:39about the AI supply chain and kind of an 6:41sbom or software build materials that is 6:45part of the solution where are those 6:47python packages being downloaded from 6:49you know how is who rolled out my AI as 6:52a service platform and who configured it 6:55and who enabled logging and really 6:57helping customers to challenge 6:59assumptions around all of that kind of 7:02backend development production 7:04environment and flow so that way in the 7:07future when you know another uh AI 7:10supply chain attack happens and some 7:12python package got back doored for 7:14example customers can quickly say okay I 7:17know exactly what I used in my solution 7:19and I don't have to worry about it 7:20obviously AI is still a very new field 7:23so I I need you to be honest with me on 7:25this one from your POV do we actually 7:28have the people or the skills that we 7:30currently need in order to properly 7:32secure AI like is there a skills Gap 7:35somewhere there's a massive skills Gap 7:37if you went on LinkedIn three years ago 7:41um the amount of people that said they 7:43were data scientists or AI Security 7:45Experts versus three months ago is night 7:48and day right everyone's claiming to to 7:51be an expert in this field or they're 7:53rushing to to train up and and the 7:57reality is there's just a massive skill 7:58shortage there's already huge skill 8:00shortage and security already and now 8:03this is a new you know brand new space 8:05that we need to be constantly you know 8:07upscaling our our both our internal 8:10security staff but also you know making 8:12developers and data scientists and 8:15management aware of those those 8:17different security risks and and making 8:20sure that you know we're we're 8:21partnering with the right people that 8:23come help test or we building out that 8:25skill set in house I'm I'm trying to 8:27figure out whose responsibility It 8:29ultimately is though like is it the 8:31developers responsibility to be thinking 8:33about security as they go about building 8:36their projects or is it somebody else's 8:38yeah great question I think it comes 8:40down to a shared responsibility model 8:43and having those discussions with your 8:45vendors and your different developers 8:47and your data scientists you know a lot 8:49of people that are training and tuning 8:51these models they say that the security 8:55of the model is not their responsibility 8:57and you know I liken that a lot to you 8:59know a web application developer saying 9:01I don't need to code an application 9:03securely just put a web application 9:05firewall in front of it right it's awaf 9:08is not a good replacement for secure 9:11code just as as a you know AI firewall 9:14or guard rails are not you know a good 9:17solution for securing models so it's a 9:21shared responsibility for whoever is 9:24training and tuning those models or 9:25supplying those models it's shared 9:28responsibility with the develop Vel 9:29opers and the production teams to build 9:31out you know decent guard rails so you 9:34know they might not prevent every attack 9:35but you can at least you know have a 9:37canary in the coal mine and know if your 9:39applications attempting to be abused or 9:41whatnot or your Solutions being abused 9:44and then it's you know having those 9:45conversations with your your AI as a 9:48service vendors like you know Azure ml 9:51or Watson X or AWS sagemaker you know 9:54what testing has gone into this platform 9:57how do I show Providence of the models 9:59and and the the AI supply chain and that 10:02you know that esbon how do I know what 10:05testing you've done on this platform or 10:07this this application that you've built 10:09for me and what expectations do you have 10:12of myself for testing my my application 10:14as it goes live since you are a hacker 10:17and I'm assuming that you enjoy your job 10:20do you get excited when you find a gap 10:23absolutely um it's what we live for as 10:26much as we love you know breaking stuff 10:28and finding apps and and circumventing 10:31all those controls it's also you know as 10:33you are in the industry for a while you 10:36kind of get tired of only burning things 10:39down and so starting to have those 10:41relationships with the blue team and 10:43with the developers and saying hey we're 10:45tired of finding this bug in every 10:47single application you create you know 10:49let's take a look earlier on in the 10:51development life cycle let's let's solve 10:54this from the start so it doesn't make 10:56its way into every app these are the 10:57controls you should put in place or 10:59these the detections you should have and 11:01if if those primary controls fail here's 11:04how you should be you know approaching 11:06this well then what's your favorite 11:08thing that you've ever hacked Chris 11:10definitely some breaking into some uh 11:13military bases back in the day was was 11:16was a lot of fun we've haded major major 11:19sporting events um most of the big Banks 11:22and and you know exchanges in the world 11:26uh nuclear power plants defense 11:28contractors chemical 11:30manufacturers it's it's uh Telos uh law 11:34enforcement intercept systems everything 11:37you name it we've we've hacked it and 11:39I've had the pleasure to lead you know 11:41the majority of of uh big engagements 11:44well Chris thank you so much for joining 11:47us today it's funny I feel like I felt 11:49my anxiety Spike some when you were 11:51talking but then I also felt relieved at 11:53the very end when I like oh but that's 11:55okay he's one of us he's one of the good 11:56guys so um we really appreciate you 11:58thank you for that Insight I feel like 12:00we could really just talk about hacking 12:02all day so now knowing what we're 12:04dealing with in terms of risks let's 12:06find out how you deal with it from aws's 12:08Resident security expert Mita Saha 12:11welcome to the show thank you Albert so 12:14let's get straight to it how has gen aai 12:17changed your job it has definitely 12:20revolutionized the whole security 12:22landscape it has introduced both 12:25challenges and opportunities 12:28specifically in my job I know that now I 12:32first need to understand how gen works 12:35so that I can understand how can I 12:37secure it we also need to make sure 12:40given as an architect when I'm trying to 12:42design new Solutions or trying to do a 12:45migration I need to be more conscious 12:49about the tools I I'm using am I able to 12:52leverate geni for an efficiency in my 12:56current work that is one landscape where 12:58we are trying to learn gen learn the 13:02benefits of gen and leverage it so it 13:05has definitely impacted I would say in a 13:08more positive and exciting way because 13:11now I'm trying to learn new things and 13:14implement it as the same time can you 13:16give me an example or two about the kind 13:18of challenges that gen can pose you know 13:21through the hands of Bad actors for you 13:23but then also how you've been able to 13:25mobilize the power of gen gen has 13:28revolutionized to hold hacking landscape 13:30like on one hand it introduced new 13:32attack vectors and Amplified the 13:35existing threats generating more 13:37convincing fishing emails generating 13:40more def fake audios or even crafty 13:44evasive Mals on the other hand AI power 13:48tools can automate and streamline 13:51vulnerability detection vulnerabilities 13:54has always been there in our security 13:55landscape rather in the industry be it 13:58any industry 13:59so we can now leverage the AI power 14:02tools and it can help us automate and 14:04streamline the vulnerability detection 14:07process it can help us enable faster and 14:09more comprehensive security assessments 14:12we cyber Security Professionals must 14:14adapt rapidly employing AI for defense 14:18while simultaneously we have to mitigate 14:21the risks possessed by the AI driven 14:24threats what's the most common mistake 14:27that people make when they're securing 14:29their data or their AI as I have been 14:32fortunate to also write a white paper 14:34have been part of a white paper we have 14:37collaborated uh featured a few months 14:39ago on securing gen and we have stated 14:42in our white paper that only 24% of the 14:46current gen projects have a security 14:49component in it even though 14:5381% of Executives say secure and 14:57trustworthy AI is essential as you can 14:59tell this suggests that many 15:01organizations are not prioritizing 15:03security for their AI initiatives from 15:06the start that is a potential oversight 15:09like nearly um 70% of Executives say 15:12Innovation takes precedence over 15:14security when it comes to gen now 15:17deprioritizing Security in favor of 15:19innovation could lead to some 15:21vulnerabilities the number one mistake 15:23that we can surely fall into is 15:26neglecting the security fundamentals or 15:29having an immature security culture in 15:32our organization which can leave our 15:35organizations ill prepared to address 15:37the conventional threats like malwares 15:41social engineering that take new forms 15:44with chaii you talked about prioritizing 15:48Innovation over security so with that 15:51prioritization let's say that some of 15:53the companies that are listening to us 15:55right now um either did do that right 15:58they've rushed to get AI ready and they 16:00bypassed the security they thought you 16:01know what we'll just do this later what 16:03can they do now what's the first thing 16:06that you would advise that they do in 16:08order to in order to Rebound in order to 16:10rectify that situation one immediate 16:13step that we can take is to conduct a 16:15comprehensive security audit risk 16:18assessment best in my opinion will be to 16:21pause the deployment right now and then 16:24conduct the security assessment instead 16:26of doing them in parallel but if in some 16:29organization in some business if that is 16:31not a possibility then at least just run 16:34the security assessment in parallel to 16:36whatever you're doing right now this 16:38audit should be performed by a team of 16:41cyber Security Experts can be from your 16:43own organization or you can hire from 16:45any of the other cyber security forms uh 16:48that you trust in and who can identify 16:52the vulnerabilities the potential attack 16:55vectors the security gaps and the a area 16:59of 17:00non-compliance with your industry 17:02standards and the AI initiative that 17:05your organization have taken second will 17:07be conducting the right threat modeling 17:09to uncover the security gaps in your AI 17:13environment and to determine how the 17:16policies and controls need to be updated 17:18to address those New Gen threats then 17:21conducting right kind of penetration 17:24testing to simulate those attacks and to 17:27identify the potten itial 17:29vulnerabilities that can be exploited 17:31and then finally I would say evaluating 17:33the systems to understand how the data 17:35is handled and the data handling 17:37practices like Access Control encryption 17:41mechanism and other potential points 17:43like networking and the most important 17:47part of this assessment should be a 17:49detailed report as a security um 17:51engineer consultant myself I believe 17:55that if I just come to my customer or my 17:58partner with the problems that won't 18:00help them right so I think a a key 18:05output of this assessment should be a 18:07detailed report I should be able to 18:09guide or recommend I'm not saying you to 18:12do this and that but I can give you a 18:14recommendation if I'm doing an 18:15assessment in your experience what's the 18:18number one potential AI related attack 18:21that's most likely right now and how can 18:24that be prevented if I have to call out 18:26one top AI related attack that concerns 18:30our SE Suite which which I have read on 18:33is uh adversarial attacks on the AI 18:36systems these attacks involve 18:38manipulating the input data to an AI 18:41model to deceive the model in a way that 18:45it causes the model to make incorrect 18:48predictions or 18:50classifications and generate harmful 18:52outputs these attacks exploit the trust 18:56and Reliance we have played based on the 18:59AI systems and it can have severe 19:02implications on the security and 19:05integrity of the systems that 19:07organizations depend on so this is going 19:09to be a rapid fire round where I'm going 19:13to list some different scenarios objects 19:18I guess and I'm going to ask you just 19:20for a simple yes or no for each of these 19:23as I call them on out using geni hackers 19:25can compromise the following your phone 19:28phone yes your TV oh yes your it system 19:34yes oh your car yes depends how how 19:39advanced I have uh built my car but yes 19:43oh well me this is a lot okay your home 19:45internet kind of um I am trying to say 19:49no but again if I don't set up the right 19:52measures then yes okay you're 19:54refrigerator not right now but again yes 19:58if I is using a iot the iot version of 20:01my refrigerator and finally your dog 20:06oh you know what yes probably 20:11yes and I can give you a use kiss for 20:13all of them but yeah how can J AI hack 20:17your dog so de fake audios say I'm 20:19working and I am not always home with my 20:23dog and I have a device through which I 20:26can talk to my dog when I'm not at home 20:29now if that device is compromised and 20:32compromised with de fake audios the J 20:35bot can actually fake my exact voice 20:39just not my dog even my friend my mother 20:41who knows me very well right and if my 20:44mother gets a call not from my phone 20:46number because my phone is with myself 20:49so the J I bought will not have my phone 20:51number but say if I'm calling from a 20:53different device and that is the Deep 20:56fake audio they do and if my coming back 20:59to the dog's reference if I am saying 21:01something to my dog and my dog is very 21:04well behaved listens to me and I tell 21:07him or her to do something in the house 21:10which can become very scary I can just 21:12say just go to the backyard and go to uh 21:15a neighbor's house and get out of the 21:16house so deep fake audios is definitely 21:19something to look out for so that's why 21:22we should be always 21:24evaluating whoever is calling us or how 21:27it is happening so okay well we're 21:30relieving this feeling a lot more 21:32confident I 21:34guess that was I think that you've given 21:37me so much to to chew upon right here so 21:39I thank you so much and in fact Mamita I 21:42appreciate all the time that you shared 21:44with us today this has been beyond 21:46insightful I'm going to speak on behalf 21:48of the listeners and viewers if I may 21:50and say like whoa you know you you've 21:52given us a lot to look out for but then 21:54also you've given us some great 21:56confidence and faith in what can be done 21:59utilizing geni appreciate you for that 22:02now again I also want to give a shout 22:03out to Chris thank you Chris for joining 22:05on today's podcast and friends that's it 22:08for today so thank you all for listening 22:10thank you for watching and of course if 22:12you have thoughts please post them in 22:14the comments below and I promise we'll 22:16see you again soon 22:19[Music]