Practical AI Security Explained: LLMs, Agentic AI & AI Attacks

0:00So, I'll be telling you some foundations

0:02for AI security, right? So, in order to

0:05understand the attack surfaces, you will

0:07need to know what is the types of AI

0:10that we have here. Okay? We have been

0:12hearing this term now constantly. It has

0:14almost become like a buzzword over here.

0:16Things like generative AI, agentic AI.

0:19Okay? But, what do they mean? What are

0:20the differences between them? Okay? What

0:23are your different different types of AI

0:25that we have? What are the different

0:26types of AI models that we have? What

0:27are the applications? Okay? So, that

0:29we'll see in the foundations part. And

0:31here we will even understand the attack

0:33surface. So, for that we'll have to

0:35understand what are the components of

0:37building an AI system. Okay?

0:40And then how those components can be

0:42exploited. Okay? So, those components

0:44itself will act as the attack surface

0:46there. Okay? Then how they're exploited,

0:49that I will tell in the second section

0:51that we have where you can exploit the

0:53common AI and ML by ML we mean machine

0:57learning machine learning

0:58vulnerabilities over here. Okay? So, how

1:00ML LLMs, right? Your large language

1:03models, agentic AI systems, how they get

1:06attacked over there.

1:07Okay?

1:08Then

1:10we'll see the defense part that how do

1:13we defend these AI applications and AI

1:16systems here.

1:17Okay? So, what are all the security

1:19controls for both for your machine

1:21learning pipelines and for your LLM

1:23pipelines also we will see here what all

1:25security controls are needed.

1:28Okay? Then how governance and compliance

1:31is linked to AI security, that part also

1:33we will see here towards the end.

1:36Right? Then after that I will

1:38tell you

1:40that what are the learning pathways over

1:42here that how you can begin with AI

1:45security and what also this particular

1:48session will act as

1:52a stepping stone, a guiding light only

1:54for you that

1:55where can you prepare

1:57this AI security from? Okay, so these

2:00topics, whatever you'll see in the first

2:02four sections, are the ones that you

2:05have to focus on while preparing for AI

2:08security, right? AI security-related

2:11jobs that are there, okay?

2:14So, yeah, and towards the end the final

2:16thing, we will have a Q&A session, this

2:19way.

2:20Right? But, you can feel free to ask

2:22your doubts in the in between also, over

2:26here. Right? If it is something that

2:28will be react-

2:31involved later, that will be coming up

2:33later, so I will let you know

2:35that particular doubt, okay?

2:38So, let's begin with this. So, first

2:41thing first here,

2:43what is AI, right? And

2:46why are these AI systems and AI

2:48applications different over here?

2:53Okay? So,

2:55when we see when we talk about AI,

2:57you'll see there are so many different

3:00terms over here. Things like AI, machine

3:03learning, deep learning, generative AI,

3:07agentic AI, predictive AI, computer

3:09vision, natural language processing, so

3:12many things are there here.

3:15Right? But, what is this? Why is so many

3:19terms? Why all this confusion here?

3:25Okay, so that is what I will clear

3:26clarify here.

3:29And we begin with this, that is AI,

3:32right? The most commonly used term that

3:34we see here is artificial intelligence.

3:37Analyze the large data sets to predict

3:40the output over here. AI in the end is

3:44the currently out right, right?

3:47We know it in this machine learning form

3:49of AI, right? So, machine learning is

3:51one form of AI here.

3:53Where we teach machines how to learn,

3:56okay? And learn from where? From data,

3:59okay?

4:00Machine learning helps us to find the

4:03patterns in the data.

4:05Okay? So, say if you are using AI for

4:08cyber security, okay? And you are

4:11building one AI-based security control

4:15that can detect network attacks for you.

4:18So, we are also so as we are growing up

4:20like in this world, so we have we also

4:23have our brain, right? So, our brain is

4:27helping us to find the patterns in the

4:29data. And see, we are gathering data on

4:31a constant basis. See, we have our sense

4:34organs, isn't it? And sense organs like

4:38so whatever we are seeing, whatever we

4:40are hearing, right? So, all those

4:42signals are going into our brain and our

4:45brain is finding out some of the

4:47patterns out there and depending on our

4:50past experiences, we react that way.

4:55Isn't it? So, with experience, right? We

4:57learn also, right? And similarly, these

4:59AI systems also so one more thing is

5:01there, they can also learn from the new

5:04incoming data.

5:05Is there, right? So, these systems have

5:08the capability of learning over here.

5:11Right? So,

5:13you had AI-based systems before also,

5:16right? Before this machine learning. So,

5:17in fact, AI is not a new field of study.

5:21So, in 1955, there was actually this

5:24invitation. So, John McCarthy, one of

5:27the researchers, they had invited lots

5:29of other researchers also

5:31for some 10 weeks of program like it was

5:34a kind of summer school held in

5:35Dartmouth over there. So, there they

5:38coined that this word AI that okay, that

5:40that this is a field that is talking

5:42about making systems that show this

5:46intelligent behavior over here, this

5:48thing. Okay, definitely the meaning of

5:51this intelligent and intelligence, these

5:53things they change, right? So, we

5:56see most of the systems that we

5:59relate to be human-like, we think it is

6:02AI over there, right? So, as soon as we

6:04see a robot, we think that this is like

6:07AI, artificial intelligence. So, it's

6:09not necessary that

6:11within it only that the machine learning

6:14is being used, right? That is not

6:15necessary, right? So, you could have

6:17rule-based AI also, right? And

6:20initially, right? It till like before

6:22your 1980s over there, you were using

6:25the rule-based AI. Okay, simple. So, if

6:28you have some idea about programming, so

6:30if-else logic that is there that if So,

6:32as if this is the case, then do this,

6:34else then do this. Okay. So, based on

6:37this itself, you were trying to create

6:38AI systems over here, right? In fact,

6:41initially, we had some of the systems

6:43called as this expert systems. Okay.

6:47They were maintained by domain-level

6:49experts. Okay. So,

6:52you had to say one of the example of

6:54this expert system is a chatbot, right?

6:58Called as Eliza over here. Okay. And

7:02what was happening in this expert system

7:05was it was

7:07trained just to answer questions like a

7:09therapist would.

7:11Okay, psychologists are there, right?

7:13They are like their clients, they ask

7:15them the questions and the the therapist

7:17will give them the answers or solutions

7:19accordingly, right? Or they may ask them

7:21a counter question, this way.

7:27Right? Okay. So, if you want to create

7:32a security control, AI-based security

7:34control, say AI-based network attack

7:37detector, so what kind of data will you

7:40require for that? What kind of

7:42historical data would you need?

7:45Can we train a model that can find

7:47patterns in our logs? So, say in SOC we

7:51are using these same tools out here,

7:53right? Like this Splunk, Elk Stack.

7:57Okay, Sentinel.

7:58Right? So, there all the logs are

8:01aggregated there in one place.

8:03Isn't it?

8:07So, what we can do, we can collect all

8:11that attack-related data in your normal

8:13network traffic and train a model to

8:16identify patterns in that. Historical

8:19logs, network traffic.

8:22If you have a collection for that, from

8:25that you can train a model to identify

8:28the patterns. Okay, what is the pattern

8:30of attack over here? What is the pattern

8:33of normal traffic here?

8:35Okay? So, for that we are utilizing this

8:38machine learning. Okay? So, so see AI is

8:41a broad field where we are building

8:43systems that is kind of mimicking this

8:45intelligent behavior. This one.

8:48Okay? So, then

8:50machine learning is that field of AI

8:53where you are creating systems that can

8:57learn from data by recognizing the

9:00patterns, by doing some pattern

9:02recognition there. Okay? And deep

9:05learning, see, deep learning is also

9:07machine learning only, right? So, I will

9:09say that deep learning is a further

9:11subset of machine learning. Okay, so see

9:14if AI is the like the outermost part of

9:16your Venn diagram, machine learning will

9:18be there within it. Okay, then deep

9:20learning will be there and so on here.

9:22Okay? So, you can see deep learning is

9:25machine learning only, but how it is

9:27different from traditional machine

9:29learning is that deep learning will use

9:31this neural networks here.

9:38Okay, and what are these neural

9:39networks?

9:40They are

9:42networks that are made out of these

9:44artificial neurons here. Okay, so

9:47actually the technical term for this

9:48artificial neuron is this perceptron

9:50that we have here.

9:52Okay, perceptron is the term that is

9:53used for this and

9:56here

9:57just like how our human brain, it is

10:00made up of

10:03these brain cells, isn't it? Brain cells

10:05that we call as neurons. Okay, so the

10:08researchers they were inspired by the

10:10our brain itself that how our brain is

10:13learning from incoming data. Okay, how

10:15it is

10:16because our brain is also very

10:17sophisticated pattern recognition

10:21tool itself that we have.

10:24Right? And not not just that it it it it

10:27it can make decisions also, right? Our

10:29brain is helping us to make the

10:30decisions out there, right? And our goal

10:32over here is to create intelligent

10:35agents that can make the decisions also.

10:40Okay, so to improve this pattern

10:43recognition technique, right? Which was

10:45there in machine learning, the

10:46researchers developed these neural

10:48networks. Okay, and when neural networks

10:50are used, we call it as deep learning

10:53here.

10:55Okay, so

10:57here

10:59as you can see, this perceptron is just

11:01one neuron. So our entire brain is not

11:04just made up of one neuron, right? There

11:06you will see there will be like a

11:07combination of neurons, right? In fact,

11:10an entire network of these neurons will

11:13be there here. So that is what is

11:15happening in deep learning also, right?

11:18We are

11:20utilizing these

11:22multiple layers of these artificial

11:26neurons which we're calling as

11:27perceptron.

11:28Okay, which we this entire structure

11:31what I'm drawing over here which I can

11:32call it as a multi-layer perceptron

11:35also. Yeah.

11:38Okay, so this we can say is a deep

11:40learning algorithm that we have here.

11:43Okay, so you know now the difference

11:45between like what do we mean by AI, what

11:48is machine learning, what is deep

11:49learning, right? Now

11:53we have one very specific type of deep

11:55learning here, right?

11:57One

11:59architecture was there that is called as

12:01this transformer architecture.

12:03Right? Now with this transformer

12:06architecture, right? What what was

12:08actually happening? You had some kind of

12:10neural networks and the researchers were

12:12trying to improve upon that neural

12:14network only out there. Okay, actually

12:17there was this use case as you know, see

12:19so Alexa was also making use of AI. Yes,

12:22Tesla self-driving cars, that is also

12:25using AI. In fact, Grammarly is also

12:27using AI over there.

12:29Right? See,

12:31when ChatGPT came out over here,

12:33Grammarly became very explicit in saying

12:35that yeah, we are using AI, we are using

12:37AI, but before that also they were using

12:39AI though they were not marketing it

12:41that much. Yeah, because there was no

12:42hype around it.

12:43There.

12:45Okay, so

12:46see it is doing sentence completion,

12:48Grammarly. It is doing sentence

12:50prediction out there, right? So that is

12:52predicting a better sentence for you.

12:54Okay, it is completing sentence. Okay,

12:57some of the applications where you're

12:59doing language translation. Okay, so all

13:02these are use cases of deep learning

13:04only here. Whatever I have mentioned,

13:06deep learning and machine learning,

13:07right? YouTube is also using AI, okay,

13:09yes for all the recommendations, okay?

13:13So

13:14you are doing

13:16see, your feed will be different, my

13:18feed will be different over here, right?

13:20Why? Because YouTube is recommending us

13:23different videos. Okay, so that is also

13:26happening due to one of the technique

13:28called as unsupervised learning.

13:31Right? Profiling, yes, profiling is also

13:34a example is a use case of unsupervised

13:37learning also here. Yes, music apps like

13:39JioSaavn, they are also So, they are

13:41also using the same kind of profiling.

13:43The end user profiling over there based

13:45on them that based on the end users

13:47choices, they are recommending the

13:49further videos, music, and so on here.

13:52Okay, yes, correct spellings over here

13:55that is also a use case of AI which is

13:57using this deep learning models over

13:59here.

14:00Okay, Google Assistant, yeah, like so

14:02all these Siri, Alexa, your Okay,

14:04Google, right?

14:06Perfect, perfect. So, we have all these

14:09AI applications also over here. Okay, so

14:13traditionally, we were we were calling

14:16all these things as predictive AI,

14:18right? And see, you can see how these

14:20fields are relating to each other. So,

14:23as you know,

14:24AI, okay, AI and data, the field of

14:27data, they are closely related. As I

14:29told you, machine learning is happening

14:31why how? It it is finding out patterns

14:34in the data. Okay, so you are doing some

14:36kind of data collection. Okay, so when

14:39you see the machine learning life cycle,

14:42when you're building these machine

14:44learning models, you always begin with

14:46data collection. Okay, so you're

14:48collecting your data, but your data

14:50might be in a text format, it might not

14:52be structured properly, right? So, you

14:54have to properly structure it, you have

14:56to clean it, you have to process it,

14:58right? You have to transform it, right?

15:00So, for all that thing, data science is

15:02also involved. That's why you can see

15:04this data science and

15:06artificial intelligence, they are two

15:08closely related fields that we have

15:10here.

15:11Okay, and then I told you like not all

15:13AI was using machine learning only,

15:15right? The example of expert systems I

15:18told you before in the beginning. Okay,

15:20so that's why you can say that Okay,

15:22when we are using machine learning right

15:23now that is another subset of AI then

15:26now when you're using these neural

15:27networks and then you have these deep

15:29learning. So these are further subsets

15:32over here. Okay, and this large language

15:35model like these LLMs like these GPT

15:38models that we have your cloud models

15:40different different cloud models

15:41different different Gemini models. So

15:43many open source models are there. Okay,

15:46so they are all examples of these large

15:48language models here. Okay. So they are

15:51nothing but deep neural networks only.

15:54It's just that they are utilizing some

15:57a separate kind of architecture that we

15:59call as transformer architecture.

16:02Okay, so David is a type of deep

16:05learning only. Okay, but it is a very

16:07specific time of type of deep learning

16:09utilizing this transformer architecture.

16:11That is why we are saying this LLM is a

16:13further subset of deep learning.

16:17Okay, and you see another field also

16:20here this NLP. This stands for natural

16:23language processing. Okay, so whenever

16:26we are building data whenever we are

16:28building models and where you are

16:30dealing with text data. So they are

16:32using NLP is a kind of no-brainer. You

16:34will have to use it over here. Okay, so

16:38with NLP you're actually trying to teach

16:40these machines the meaning of the words

16:43the meaning of words, grammar, sentences

16:46that way. Okay, so many of the examples

16:48that you also told me things like these

16:52sentence completion out here,

16:53correction. Okay, the Grammarly also for

16:56example. So they are utilizing these

17:00natural language processing out here.

17:02Okay, and of course these large language

17:04models like your GPT models and cloud

17:06models. They they cannot be built

17:08without using this natural language

17:10processing. Okay, so that's why I'm

17:12telling you that this field of study is

17:14also intersecting with AI over here.

17:16Okay, and if you want to add more

17:18things, I can add another category over

17:20here. Right? And I can

17:25call it as

17:27CV over here. And that is your computer

17:30vision. Okay? So, you want We have seen

17:32that our AI models, they have the

17:34ability to process images also, videos

17:37also. They can You have robots where

17:39you're putting in the sensors over there

17:40and they can capture the visual inputs

17:43and within those robots, they will be

17:45the models.

17:47Right? AI models that can process those

17:51images. Okay? And then when we take the

17:53case of self-driving cars also. Okay?

17:56So,

17:58in self-driving cars,

18:00they also need to see their

18:02surroundings, right? They need to see

18:03which all cars are there in front of

18:05them. They need to see the

18:06pedestrians, the objects, right? So, for

18:08that also,

18:10in computer vision will be involved

18:12there also.

18:13In that case, right? So, you see with

18:16and I see it it it it is all coming

18:18together, right? When we said that what

18:19is AI initially, right? We all said

18:22that, "Okay, we are trying to create

18:23systems that are

18:27just like human, human-like in

18:28intelligence out there.

18:30Right? That can mimic the human

18:31intelligence this way. Right? So,

18:33humans, like we are able to think and do

18:36that way because we can see, we can talk

18:39all these things, right? So, the ability

18:41to see, to understand language, to talk,

18:44right? So, that can be given through

18:46this computer vision and natural

18:48language processing also. Here.

18:51Okay?

18:53Now,

18:54yeah.

18:55This These three terms here.

18:58Predictive AI,

19:00generative AI,

19:02agentic AI.

19:06Okay?

19:07So far, whatever I've told you, these

19:10use cases, everything

19:12based on your data,

19:14based on your data over here when your

19:17AI models are looking at the patterns in

19:20the data over here.

19:23Right? When they're looking at the

19:25patterns in your data,

19:28based on that they are making some kind

19:30of prediction only.

19:32See, if you're talking about Netflix

19:34recommendation, that is also predicting

19:38what

19:40the user would like to see.

19:43Okay, same for the music.

19:45Same for sentence completion. It is

19:47predicting the next word.

19:50Same for language translation. It is

19:52predicting what can be the translation

19:54of that particular

19:55sentence out there.

19:57If you look at other use cases also

20:00here.

20:06That

20:09you are using AI as a security control.

20:12So, in that case also, what? Based on

20:14the training, based on the patterns,

20:17based on whatever it has learned during

20:19training, right? Once you have taught it

20:21what are the patterns of a network

20:22attack and what are the patterns of a

20:24normal traffic. And then after that,

20:26when you're deploying these models,

20:30so thereafter,

20:34the model is also making a prediction

20:36based on your incoming data, based on

20:38your incoming packets, based on the logs

20:40that it is seeing. It is predicting

20:42whether

20:43it is normal network traffic or whether

20:45it is network attack.

20:47Okay? If you're giving it a image,

20:49right? So, yeah, if you are seeing these

20:52image detectors only out there. Okay?

20:54Which we said they're like using this

20:55computer vision here. Okay? So,

20:59there we can train models to detect

21:01images. Okay? So, there is this Google

21:03Lens that you're using in your phones.

21:06Okay? Where you just take the image and

21:08the AI there will predict which

21:11category it belongs to. Okay, if you

21:12take the image of a bird, it will even

21:13tell you which bird this is. Okay?

21:17So, there also it is doing predictions.

21:19That is also predictive AI.

21:21Okay, if you're predicting the stock

21:24prices, AI your quant teams there also

21:26building these AI models, right? That

21:28can

21:28suggest you when to

21:31purchase a stock, when to sell a

21:33particular stock out there, right? So,

21:34that is also

21:36predictive AI only.

21:39Okay? So, then you have this generative

21:42AI and all this ChatGPT revolution when

21:46it came up November 2022, right? So,

21:48that is your generative AI. Why? Because

21:52it is being used to create some content.

21:55Okay? If you tell it to create some code

21:58to do the coding, right? That is also

22:00like it is generating the code, isn't

22:02it? And then

22:04if you ask it to generate an image,

22:05right? So, again generating generative

22:08AI, right? So, idea is to generate some

22:10kind of thing.

22:11Okay? But when you go when you dive deep

22:14into its inner workings there, that how

22:16it is working, you will see that

22:18everything is predictive AI only.

22:21Why? Because even in generative AI, even

22:24these large language models like these

22:26GPT models and Claude models and Gemini

22:28models and other open-source models that

22:30you're using,

22:33it is just predicting the next token for

22:36you.

22:38Okay?

22:39Token by token I mean the next word

22:41here. Okay? Technically when you will

22:44see like tokens can be like combination

22:45of a word and punctuation marks, like

22:48your combination of word and spaces or

22:51it if you have a very long word, a word

22:53can be broken down into multiple tokens

22:55also over there. This way.

22:57Okay? So,

22:59your generative AI systems like this

23:02large language models

23:04are

23:05predicting the next token itself for

23:08you.

23:10Okay, it's just that we are using

23:12predictive AI, using the prediction

23:14power of AI, we have now even taught

23:18these machines how to generate data

23:21here.

23:30>> [snorts]

23:31>> Okay? And you have this concept of

23:35agentic AI also here.

23:38Okay? Though this term agentic AI, see,

23:40generative AI is not a very new term. It

23:44was being used, right? Though not

23:48see, day-to-day and like a buzzword and

23:50common word like a common household

23:52word, this GenAI has become now, right?

23:54After this LLM large language model

23:56revolution after ChatGPT was launched.

23:58GenAI became a very common term over

24:00here. Okay? But in your academic spaces,

24:03right? In your academia, this GenAI was

24:06being used before also.

24:08Right? In fact,

24:11using AI for sentence completion, there

24:13also you're generating a kind of

24:14sentence only over there, right?

24:17Isn't it?

24:18So,

24:20yeah, you had some kind of models also

24:22called as these generative adversarial

24:24networks, right? Which was launched in

24:262014 only, right? So,

24:30you could use these GANs for creating

24:33deepfakes even before you could create

24:36these deepfakes over here.

24:40Okay?

24:41Before these LLMs, the problem for

24:43deepfakes were already there. Okay? You

24:45had some use cases where you could use

24:48AI for enhancing these images, enhancing

24:51the videos, right? You must have seen

24:53some

24:55movies, right? The old movie will be

24:56black and white, but you must have seen

24:58that okay, the enhanced version of

25:01so-and-so movie over here in color also

25:03it is there. Okay, so they were all

25:05built using these GenAI tools like this

25:07generative adversarial networks also.

25:09Yeah.

25:12Okay? So, then next we have this concept

25:16of agentic AI. Okay? So, though this

25:18agentic AI is a term that is newly

25:21coined, right? It was coined in 2024

25:23only. This one. But the term agents is

25:27not new.

25:29Okay? Agents is a term that is used

25:35very frequently and now and earlier also

25:39it was used this term agent over here.

25:42Okay? So, agent is not a new thing here.

25:47Now,

25:48say we have two words here.

25:51Okay? When we talk about GenAI, we think

25:53about a chatbot over here that can

25:55answer questions for us. Okay? So, you

25:58have this term chatbot.

26:00Okay? And you have this term

26:05agent over here. These two terms,

26:07chatbot and agent.

26:09What do you think is the difference

26:11between these two terms here?

26:14So, Soham says chatbot will suggest and

26:16give the idea and code, but agents will

26:19perform it. Yes. So, so what we have a

26:23chatbot just has It's just a brain on

26:26its own this way.

26:28Okay? So,

26:30it has no tools, it has no external

26:33tools to perform some task out here.

26:36Okay? So, just imagine if you're giving

26:39a chatbot and we say that the chatbot

26:41becomes an agent, it can be more

26:43intelligent out here.

26:45See, it the intelligence aspect is

26:47already there, right? Because chatbot is

26:49a brain this way. But it's a it is just

26:51a standalone system. Say, for example,

26:54when

26:56ChatGPT was first launched, okay? So, it

27:00did not even have that web access tool,

27:02the web search tool. Right? So, if you

27:05would ask it questions,

27:08uh so, it had a training cut-off date,

27:09right? Its training cut-off date was

27:11November 20 October 2022 there

27:14initially. I'm talking about initially,

27:15the GPT 3.5 version when it came out

27:18over there. Okay? So, in 2023, if you

27:22ask if you would ask it some questions

27:24related to the latest news, it wouldn't

27:26be able to give the answer, or it would

27:27just hallucinate the answer over there.

27:32Okay? So, but say if you give it some

27:35particular tool over there with which it

27:37can do some web search also.

27:41Right? Then, we can say that now it is

27:43not a plain chatbot, it is a agent also

27:46that can utilize the tool over here,

27:48okay? So, the chatbot or the model, see,

27:52chatbot, it will make use of

27:55uh AI model over here.

27:59Okay? And this AI model that we have,

28:02this AI model is acting as the brain

28:05over here.

28:07So, they are agents because you can not

28:10only chat with them, you can chat with

28:13them, that is there. You can give you a

28:14questions to your answers.

28:20You your agent Siri and Alexa like this,

28:23they can set up an alarm for you.

28:25They can help you call a particular

28:27person, right? So, it means they have

28:29access to the tools, right? They have

28:31access to your calendars over here,

28:34right? They have access to your clock,

28:37okay? They have access to your

28:40contacts.

28:50>> Okay, so that is why we are saying the

28:53agent has a brain also.

28:56Right? In the form of these AI models.

29:00Okay, and these tools over here are

29:03acting as the hands of the agent so that

29:07it can perform the tasks there.

29:17Okay, so yeah, skills. So Manas, you're

29:19talking about the latest one these

29:21agentic AI systems where you're

29:23utilizing the LLMs also as the brain

29:26there. Okay, so to the LLMs when we give

29:29a complete goal over there and the

29:32complete context that what it needs to

29:35do, how it has to achieve that

29:36particular thing, what is the expected

29:38output that we need. Okay, so all that

29:41thing can be given in the form of

29:42skills, right? You can even give that in

29:44the form of a markdown file, .md file

29:47there. Right? You like your Claude

29:50agents like we can utilize them with

29:52this particular thing.

29:55Okay, so

29:57here you can see this chatbot. So this

30:00plus this tool usage here, we get the

30:03agents here. Okay, now

30:06this term agentic AI it was coined here

30:09because there was in in the academia and

30:12the industry also there was this debate.

30:14Okay, that what is an AI agent, right?

30:17Because you you had been utilizing AI

30:19agents since your '80s and '90s also

30:22over there. In fact, these expert

30:23systems were also called as agents only

30:25over there, but they were just

30:26rule-based. Okay, so there is one this

30:28very famous AI educator, right? His name

30:31is Andrew Ng, okay? Andrew Young. So

30:35he was giving a speech out there, right?

30:38And in that he said that that

30:42there are lots of kinds of intelligent

30:44systems that we have, okay? And we are

30:48talking about agents. So, agents are

30:50something that have some kind of goal,

30:53okay? They utilize these tools over

30:56here, and they use a reasoning element,

30:59which is the brain, which can be a large

31:01language model, okay? And based on this,

31:05they

31:07complete that particular task there,

31:09right? And they have the autonomy to

31:15create their

31:16own

31:19thinking over here, as in the reasoning

31:21over here, this thing.

31:23Okay, and they can improve with time

31:25also. So, they have this feedback loop

31:27also here.

31:35Okay, so agents is anything that will

31:39act autonomously, right? In order to

31:42achieve this goal over here.

31:45Right? So, for that thing, they have

31:47these guiding stones that are these

31:49brain and the tools here.

31:52Okay? So,

31:54now I will tell you just the difference

31:57between these these predictive,

31:59generative, and agentic. We have already

32:01seen the difference. I will demonstrate

32:02also practically, right? Before that,

32:04one more important thing here is just

32:07let us see all these things what we have

32:09seen

32:10in one slide only over here.

32:13Okay? So, we know that we have AI,

32:16artificial intelligence, we have machine

32:18learning, we have deep learnings over

32:21here, these things. Okay? Now, tell me,

32:24if you want to make

32:26if you want to build an AI model,

32:32okay? So, what are the minimal

32:34components? What things would you

32:36require to build an AI model?

32:38What are the requirements to build an AI

32:40model?

32:42Knowledge, okay, so knowledge you will

32:44get it from data, right? Data is one of

32:47the component and then

32:52So LLM model will be the end product.

32:56LLM model will be the end product. Okay,

32:59so say if you want to build an LLM model

33:02then apart from data, okay, so some kind

33:05of infrastructure, compute power you

33:07would require.

33:10Okay, compute computation power will be

33:12required there. So GPU, CPU, TPU.

33:15Okay, RAM. Prompt will come in once you

33:18have already built a model and then

33:21you are

33:22utilizing it there. Okay, yeah, that is

33:25another thing that you can utilize these

33:27GenAI tools and you can give a prompt

33:29and with prompt itself we can build an

33:31AI model. That is a that is a separate

33:33thing out there, but uh yeah, when when

33:36when when when a prompt also you're

33:37giving

33:38in the end if you look at first

33:40principles, you will end up with some

33:42kind of data.

33:43You will you will need the computation

33:45power or your entire infrastructure and

33:48yes, as you have rightly

33:51pointed out, it is the algorithm over

33:54here. This one. Okay, so this

33:58transformer is a kind of architecture

34:00which I will tell you about. This

34:02transformer, see over here as you can

34:04see in this slide, we have this

34:06transformer architecture.

34:08Okay, so transformer is something using

34:12which like which is acting as a

34:13blueprint for all these large language

34:15models like GPT models, Claude and

34:17Gemini models.

34:19Over here. Okay, but at the core of it

34:21if you look, to build an AI model, you

34:24require data algorithms

34:28and computation power. Okay, as I said

34:30that

34:32machine learning is about pattern

34:33recognition. Okay.

34:35But pattern recognition where?

34:38In the data.

34:39Okay. Pattern recognition who? Who is

34:42doing it? That is this algorithm. Your

34:45algorithm is recognizing the patterns

34:47there.

34:48Okay. So there is a process that we say

34:50that we we call as fitting the data into

34:53the algorithms over here.

34:54Okay. We call it fit the data into

34:57algorithms. This one.

34:59Okay. So in the end in the back end if

35:02you see these machine learning deep

35:04learning algorithms, right? Even these

35:06LLMs, right? They are nothing but

35:10a sophisticated

35:12machine that can do some kind of math

35:15mathematics.

35:16Okay. So algorithms, you will see that

35:20these are nothing but equations.

35:23Okay. You'll see various equations

35:25within it. Okay. And then equations, you

35:27know, you have coefficients, you have

35:29values where you can substitute the

35:31numbers. Okay. So within this equation

35:35itself, you will be putting these data.

35:37Okay. So that's why I said that the

35:40unstructured data and all the text

35:42related data, you will all convert it

35:45into numbers. Okay. So now finding out

35:47patterns in these numbers is will be

35:49easy. Why? So for finding the patterns,

35:51simply you will just put these numerical

35:55data into these equations, which is the

35:57algorithm. Okay. And for doing the

36:03for analyzing these equations out there,

36:05you will see that of course some kind of

36:07competition power will be required. Some

36:09kind of processors, GPUs, RAM, all these

36:11things will be required out there. Okay.

36:13So once you're doing this, then the

36:15algorithm is finding out the patterns in

36:17the data. And

36:19the algorithm will see that based on

36:21your training data, it will see that

36:23okay, what is the pattern of

36:25normal network traffic, what is the

36:27pattern of

36:28attack-related traffic, right? What is

36:30What is the pattern of a DOS attack?

36:32What is the pattern of a web application

36:34attack like your cross-site scripting,

36:36this way, right? So, all that patterns

36:39it will be studied out there, and once

36:41the algorithms has decided okay what the

36:43patterns are, right? You finally have a

36:47completely built module out here.

36:50Okay? So,

36:52if

36:54initially if we are giving the correct

36:56answers

36:57to the model, right? If we have some

36:59kind of labeled data set,

37:02then that is supervised learning. Okay?

37:06If our data set does not have labels,

37:08right? We are not telling that which

37:10category the

37:13data points they belong to, then we call

37:15it as unsupervised learning here. Okay,

37:18but in reality like we are using a

37:19combination of supervised plus

37:21unsupervised, right? And then that's why

37:23we are calling it like semi-supervised

37:25learning also over here. Okay? There is

37:28another kind of learning that initially

37:30you don't have a data set, right? So,

37:33things like self-driving car, okay? Or

37:35robotics, it is very difficult to teach

37:38a robot how to walk just by giving a

37:40data set. Isn't it? It is very difficult

37:43how to to

37:45drive a car, right? The autonomous

37:47driving with just a data set. Okay? So,

37:50when such data sets are not involved,

37:52then you call it as reinforcement

37:54learning.

37:59Okay? Many of these AI models where we

38:03see that okay this particular AI model

38:05beat the person in chess, okay? In fact,

38:07in 1997 itself, the there was this IBM a

38:11model built by IBM called as Deep Blue.

38:13So,

38:15there these kind of models they were

38:16also using reinforcement learning only

38:18here.

38:19Okay? In 2017, there was a game over

38:22here where they thought that a machine

38:23can never beat a human in this game

38:25because it had way more combinations,

38:28permutation and combinations, and way

38:29more number of moves that can be done in

38:32chess out there, like more than chess,

38:34way more than chess. Okay, but in 2017,

38:38even in Go the game of Go, AI model it

38:41beat

38:42the humans, the Go champion out there.

38:44Okay, so that again was happening with

38:46this reinforcement learning, right? So,

38:47reinforcement learning means it is

38:49learning from its environment only.

38:51Okay, so if you have a robot, right? And

38:53if it is working, if it is not able to

38:55work properly, we give it a kind of

38:56punishment point, right? Negative

38:57rewards point that we call it. Okay, if

39:00it is working properly, we give it a

39:01positive point or a

39:03positive reward points that we have.

39:06Okay, so this kind of technique is your

39:08reinforcement learning.

39:12Okay, so

39:14within machine learning also, see see

39:17this way.

39:18As I told you, to build any AI model,

39:21you require data, algorithms, and

39:23compute. Okay?

39:26Based on the algorithm that you're

39:27selecting, see, you can select a machine

39:30learning algorithm, you can select a

39:32deep learning algorithm.

39:34This way. Or what we have, we have this

39:37transformer architecture also over here.

39:42Okay? And

39:44based on data, you can see based on

39:47data, the kind of data that we have,

39:50we either have supervised learning,

39:53unsupervised learning, semi-supervised

39:54learning, or like when no data is there,

39:56then reinforcement learning. Yeah.

40:00Okay, so this way we can have supervised

40:04machine learning also, we can have

40:06supervised deep learning also. Okay, we

40:09can have unsupervised machine learning,

40:12we can have unsupervised deep learning

40:15also over here.

40:24Okay. So, now some of the use cases of

40:27machine learning we see over here. These

40:29are these regression use case, okay,

40:32classification use case,

40:35clustering, and dimensionality

40:38reduction.

40:41Okay. So, regression means when your AI

40:44model is predicting some kind of value

40:48over here.

40:52Okay. So, say if you take example, if

40:55you want to predict the stock prices,

40:57okay, or if your AI model is predicting

41:01the housing prices, okay, or if you have

41:04a AI model

41:06that can predict the average customer

41:10age group there for your products is

41:13there. Okay. So, all these use cases we

41:15can put them under the regression use

41:17case here.

41:18Okay. And this is the algorithm name

41:20that we are actually using for

41:24regression. Okay. The linear regression

41:26algorithm. Okay. As I said,

41:28uh so, you will have some kind of data

41:29set a historical data say of the stock

41:32prices itself, this one. Okay. And

41:35uh the stocks, right? The stock data

41:37[clears throat] that will be fit into

41:38this linear regression algorithm, right?

41:40And then with this using the compute

41:42power, we will build

41:45uh model, right? That is that is trained

41:47using supervised learning, right? And

41:49the algorithm used was linear

41:51regression. Okay. Like that we have

41:53classification models that are

41:55predicting

41:57category over here. Right? This is not

42:00predicting a value for you. Instead, it

42:02predicts a category. Okay. By category I

42:05mean

42:07again, say if you're building a security

42:09control, AI/ML based security control.

42:12So, for detecting

42:15whether

42:17incoming traffic is belonging to

42:22normal traffic class or malicious

42:25traffic class.

42:27Okay, so these are categories only,

42:28isn't it?

42:31Okay, we are predicting the category

42:32over here. That is why

42:34we have seen the regression use case

42:37also here and the classification use

42:39cases also here.

42:41Okay, these are just some of the

42:43examples of the algorithms that we have

42:45for classification use case.

42:50Okay, so one one of this data set data

42:55This can be used to predict whether a

42:59incoming message or the email is spam

43:02email or no, right? So, this kind of

43:04data set we can use to train a model

43:06here. Okay, so for this you can see

43:08there is one column here that is having

43:11this ham, spam, all these things, right?

43:14So, these are acting as the label. Okay,

43:16so we will let our algorithm know that

43:18okay, this particular message is normal

43:21message, okay? We will let our algorithm

43:24know that this particular message that

43:25we are seeing is a spam message here.

43:29Okay, so say if this column is present

43:32in our data then we will say it is

43:33labeled data set, okay? And if this

43:36column if I just remove it out here and

43:38if I just use this data set as it is

43:40like this then in that case it I will

43:43have to use a

43:46unsupervised learning algorithm there in

43:48that case.

43:53Okay, but

43:56in some cases

43:59we may have to use these unsupervised

44:01learning techniques also, right? We

44:03cannot do the labeling all the time.

44:07Okay?

44:08This

44:10customer profiling, okay?

44:13Next video recommendation.

44:15All this is again done through this

44:18unsupervised learning only space

44:19specifically for this by this technique

44:21called this clustering over here.

44:30Okay? What I mean by clustering, let me

44:32show it to you directly with the example

44:34only.

44:39Okay, so this is one of the program that

44:41we have made. You can generate such

44:43things using generative AI also if you

44:44know what prompts to give. Okay? So, I'm

44:47just running this first and then I'll

44:48show you the output that I'm getting

44:50here.

44:54See, along with that there is this

44:55concept of dimensionality reduction

44:58also.

45:00This one. So, this is also unsupervised

45:02learning use case.

45:05Okay? So, say

45:07when we have a data set, right? In a

45:09data set we can have complex number of

45:11dimensions also. I'll just show it to

45:12you.

45:15Okay, so this is a data set with which

45:17we can train a model to detect the

45:20cybersecurity intrusions also, the

45:21network intrusions that are taking

45:23place. Okay, so over a period of

45:25different different sessions we have

45:27collected the data. So, the number of

45:29the network packet size, okay? How many

45:32network packets are there, okay? Login

45:34attempts that were made in that

45:35particular session, what was the

45:37protocol type of the packets, okay? The

45:40session duration, encryption used, okay?

45:43IP repetition scores, okay? Number of

45:46failed logins, okay? Browser type from

45:49which the traffic was coming in, okay?

45:51Whether the traffic was coming during an

45:53unusual time, okay? And based on that we

45:56have this label column over here that we

45:59have named as attack detected. Whether

46:01attack is detected or no. This one.

46:05Okay. So, this kind of data set you will

46:07use for supervised learning, right?

46:08Where we have a supervisor, right? A

46:11trainer, a teacher, who is telling our

46:13algorithms that

46:16Okay, if so and so data points are

46:18there, okay? If this is the value of the

46:21data points, then the answer of this is

46:23one. Okay, it is belonging to the one

46:26category over here. Okay, so one here in

46:28this case is attack. Okay?

46:32For the data point that is below, it

46:34will tell that okay, yeah, that this

46:36belongs to category zero out here. Okay?

46:39So, when we are actually training the

46:40model on this data set, so then by

46:43seeing these thousands of data points,

46:45it will see the overall pattern of one

46:48that when the output is one and when the

46:50output is zero over here.

46:54Okay? And through that, your model

46:56parameters are also

46:58computed over here. Okay, model

47:00parameters are what the model learns

47:03during the training there.

47:07Okay? Once the model parameters are set,

47:10then you just have to give the incoming

47:13uh packets or the logs that are there,

47:15okay? And based on that, the model will

47:18make some predictions, right? So, during

47:20the prediction, like during the

47:22uh deployment, right? When you're

47:23testing or when you're testing it with

47:25new data also, then you don't give it

47:27the labels. Then you hide this label

47:29part there.

47:31Okay? You will just give it the incoming

47:32data. And based on that, your model is

47:35predicting whether the incoming data is

47:36belonging to category zero or one,

47:39right? So, if it is belonging to

47:40category one, then of course the alerts

47:42will be raised over there.

47:47Okay? So, this is supervised learning.

47:49Supervised learning is pretty

47:50straightforward over there and it is

47:52mostly preferred also, right? Because

47:54training a model is easier with the

47:55supervised learning out here. Okay? but

47:58because why? Because we have the correct

47:59answers that we are supplying to these

48:03models. Okay, but having such kind of

48:06label data set may not be feasible all

48:07the time.

48:09Okay?

48:10In that case, you go for this

48:12unsupervised learning here.

48:14Okay, and I told you this clustering and

48:16dimensionality reduction is one of the

48:18use case over here. So, see clustering

48:20I'm showing you. See this one.

48:26I ran this program over here and I've

48:28got this output. Let me

48:30just show you the output of this program

48:32here. Yeah.

48:34Okay? So, if you can see this

48:36what this clustering is doing?

48:39Notice, I'm giving it some network

48:42traffic there and it is doing the

48:44network traffic profiling.

48:46Okay? And it is

48:49grouped my data

48:52into various clusters, right? You can

48:54see this in in pink I have one cluster

48:56over here. This green is one cluster.

48:58Yellow is one cluster. Blue is one

49:00cluster. Purple is one cluster here.

49:01This way, right? So, different different

49:03clusters we have over here.

49:05Okay? So, see here we are doing the

49:08network traffic profiling. This is the

49:10unsupervised learning. So, see it didn't

49:12have the labels. We didn't tell that,

49:14okay, this network traffic is of which

49:16category there, right? It has just

49:18grouped them into various clusters.

49:20Okay? Then after that, what the analyst

49:22did, the analyst is just looking at

49:25these clusters over here and seeing that

49:27and matching it with the logs and then

49:29seeing that, okay, what kind of traffic

49:31it is. Okay? Then what the analyst did,

49:34the analyst has just put the label later

49:35on that, okay, yeah, this red one, this

49:38thing is DNS query. Okay, the one you

49:40have in brown, this is your normal HTTP

49:42traffic. Okay? This is if someone is

49:44doing a port scan attempt, then you are

49:46getting this uh

49:48data points in these clusters. When

49:50you're doing the file transfer, then

49:52you're getting it in this particular

49:54cluster. When you're doing the normal

49:55streaming here, then you're getting this

49:57cluster here in gray.

50:03Okay, so this is our clustering here.

50:05Okay, and this can be useful anomaly

50:07detection also. Say we have done the

50:08entire profiling for normal network

50:10traffic. So if there is some kind of

50:12zero-day attack, right? So a traditional

50:15rule-based detector, of course, the

50:17zero-day attack cannot be detected with

50:19that, right? Because zero-day means the

50:21attack for which we have no signatures,

50:23right? There is no

50:25zero-day vulnerability means even the

50:27vendor of that particular product or

50:30application doesn't know that that

50:32particular vulnerability exists, okay?

50:34So if the

50:36That only that is your people are

50:38unaware, then of course the patches and

50:41the signatures of it won't be available

50:43here.

50:44Okay, so a zero-day attacks cannot be

50:46detected using supervised learning.

50:47Supervised learning is good for known

50:50attacks, okay? But for unknown attacks,

50:53we can use these kind of clustering

50:54techniques, okay? Because and

50:57different clusters, I say if we have

50:59some anomalous cluster. So say that is

51:01forming a cluster over here in this

51:03section between the six and eight and

51:05one and two over here in X axis and six

51:08and eight in Y axis.

51:10Okay, so then we can say that that is a

51:11kind of anomaly over here.

51:14Okay, but there there is always a catch

51:16here, right? Unsupervised learning. In

51:18supervised learning, our false positive

51:20rate is very less, right? Around 5 to

51:2310% false positives only we will get in

51:27supervised learning. But in unsupervised

51:29learning,

51:30it's not necessary that any anomalous

51:33activity is malicious, isn't it? So even

51:37though we will get lots of anomalies

51:38here, but

51:41it may be lots of false positives,

51:43right? That's why when you're going for

51:44this unsupervised learning, lots of

51:46fine-tuning is required. And that's why

51:48I said ideally, when we are building

51:50AI/ML based security controls, we are

51:52using a combination of both supervised

51:54and unsupervised learning over here.

51:58Okay, dimensionality reduction. See,

52:00this is also use case of supervised

52:02learning. If you look at this data set,

52:04so okay, this particular one

52:06has, if you see number of columns here,

52:08ABC till K it has, right? So, it has

52:1311 columns here. This one. So, we can

52:16say this is 11 dimension data.

52:18Okay, sometimes our data set may have

52:2030, 40, even 100, hundreds of columns

52:23out there.

52:24Okay, so in that case, we would need to

52:27reduce the number of columns, right? So,

52:29that is our dimensionality reduction.

52:31Okay, for example, you can take here in

52:34in this case only, right? So, say if we

52:35have this X and Y axis here. Okay, so in

52:40X and Y axis, I can just represent

52:42two-dimensional data.

52:44Okay, so notice that's why I've done

52:46this PCA also, principal component

52:48analysis, okay, which is a kind of

52:50dimensionality reduction only.

52:52Okay, in in my data set here.

52:57This is my data set.

53:00Okay.

53:02I have various features like this

53:04duration,

53:05bytes sent, bytes received, okay,

53:08packets sent, okay, packets received,

53:11port. Okay, so I have six dimensions

53:14over here.

53:161 2 3 4 5 6 dimensions. Okay, so to

53:20represent this data point, I would need

53:22six dimension data.

53:24Right? Getting the point?

53:26Two-dimensional data can be represented

53:28in two two-dimension coordinate system,

53:30okay, X and Y axis. Three-dimension data

53:33can be represented in XYZ axis, okay?

53:36But we,

53:38in a physical sense, we don't have the

53:40concept of four, five dimensions and so

53:42on. Isn't it?

53:45But mathematically, multiple higher

53:47dimensions can be represented. Okay?

53:49Simply, your more number of columns here

53:52are adding to the dimensions. Okay? But

53:54this kind of data, if we want to

53:55represent it in such a graph, so here in

53:58this case we are representing those data

54:00points in two-dimensional space only. So

54:02that's why we have utilized one of the

54:06unsupervised learning techniques that

54:07you see here is this

54:09PCA over here, PCA plot that we are

54:11drawing over here, principal component

54:13analysis. Right? And here. So our data,

54:16we are break we are condensing it into

54:19just two components so that we can draw

54:21it in a graph also here. Right? So that

54:23is another use case of this unsupervised

54:25learning here.

54:27Okay? Now I'll tell you something. This

54:29is quite interesting over here. You have

54:32So of course, you have this deep

54:33learning here. And in deep learning, I

54:35already told you about the artificial

54:37neuron that deep learning you utilizes

54:38artificial neurons. And

54:41within that, I told you that one

54:44artificial neuron is called as

54:45perceptrons and multiple layer of it

54:47that is called as this multi-layer

54:49perceptrons. Okay? Now these multi-layer

54:51perceptrons only, if you have modified

54:53it so that it can view the images,

54:55process the images, if you're giving the

54:57AI the power to see, that is all

54:59happening because you're utilizing the

55:01CNN models, that is the convolutional

55:03neural networks. Okay? So this is giving

55:06it the power to see over here.

55:10Okay? And these RNN LSTM models, all

55:14these sentence translation, sentence

55:17completion use cases, right? They So you

55:20can see these RNN LSTM models, they are

55:22giving the AI the power to read over

55:24here, read or understand text also.

55:30Okay? Now in 2017

55:34the researchers, they were trying to

55:37improve this RNN LSTM models, right?

55:40They wanted to make this

55:43sentence translation, they want to

55:45improve it they wanted to improve it

55:47even further, right? They wanted to the

55:49translation to happen in real time.

55:52Let's say if you are talking with some

55:54other person in Spanish, okay? Someone

55:57other person is speaking Spanish, you

55:58are speaking English, and you want that

56:01as soon as you're speaking, they because

56:04you have some kind of AR augmented

56:06reality and all these VR glasses also

56:09and using that, the translations should

56:12happen in real time, right? There should

56:13not be much lag, okay? So, already the

56:16trans- the sentence translation was

56:18pretty good before 2017 also, but it was

56:20a bit slow, right? So, to improve that,

56:23to make it more real time, what they

56:25did, they introduced one mechanism,

56:28right? It was called as this attention

56:29mechanism.

56:31This thing.

56:32Okay? This when it was combined with

56:34your existing deep neural networks, this

56:36led to the formation of your

56:39transformers,

56:40okay? The transformer architecture.

56:43Here.

56:44Okay? And like we have sentence

56:46transformers also that can uh

56:49build the that are building all these

56:51large language models, right? That can

56:54answer your questions, okay? That can

56:57generate code, that can provide you

56:58summaries, okay? So, these are all the

57:00sentence transformers and you have the

57:01vision transformers also.

57:04Okay? Because you know now these LLMs,

57:06they have the capability to process your

57:08images also and to generate videos also

57:10over here, okay? So, with that, we have

57:12these vision transformers, okay? So,

57:14these transformers itself, as I told

57:15you, they have led to this GPT

57:17revolution that we have. All your large

57:19language models.

57:22Okay? So, we saw this predictive AI,

57:25generative AI, agentic AI also here,

57:28right? And we saw these models also,

57:31right? So, you can see so many models

57:33are there.

57:34We have machine learning based ones,

57:36deep learning based ones,

57:39and this

57:40transformer

57:42architecture, which are building your

57:44LLMs.

57:45Okay, this transformer architecture they

57:48all these LLMs they are built using a

57:50combination of unsupervised learning,

57:53supervised learning, self-supervised

57:55learning, and reinforcement learning

57:56with human feedback. So, all these

57:58learning types are applied over here to

58:00build all these GPT and all the language

58:03large language models that you're

58:04utilizing. Okay, and how agentic AI also

58:08I've covered, right? How they're using

58:10these models as a brain, right? So,

58:11earlier they used to use this So, this

58:14Alexa, Siri before these LLM, right?

58:17They were using these RNN LSTM models

58:19only as a brain of there.

58:23Okay, it's just that earlier you used

58:25you used to say, "Okay, Google." Now you

58:27say,

58:28"Hey, Gemini." over here.

58:34Okay, so now in your phones the brain

58:36like is also these Gemini models over

58:38here, these things. Okay, but it doesn't

58:40means that these new models since this

58:42this transformer architecture has come

58:44up, so we are only using this, no. Say

58:46in some edge devices, IoT devices, or

58:52where much processing power is not

58:55available, there we are still making use

58:57of these ML and deep learning models

58:59here.

59:00Okay, in many of the cases you want the

59:03proper data compliance. Okay, you want

59:06that your data, your customer PIIs they

59:09should not be

59:12read by these AI models. Okay, in that

59:14case we cannot use these transformer we

59:16cannot make use of these GPT models

59:19there in that case, right? In that case

59:20we will be preferring these deep

59:22learning and machine learning models

59:23only.

59:27Okay, so still

59:30in AI use cases, right? Be it your use

59:32case in cybersecurity also as a security

59:35control or be it as a productivity tool,

59:37right? We are utilizing

59:40all three of them. Be it machine

59:43learning, be it deep learning, be it all

59:44the transformer architecture in the form

59:46of all these GPT models out here.

59:58Okay, example of machine learning I'll

1:00:01show you over here. Simply as I told

1:00:03you, you're fitting the data algorithms

1:00:05and compute, okay? So, say I have my

1:00:09data.

1:00:11All my original data, see, I have

1:00:13collected the normal network traffic and

1:00:15DOS-related traffic out here, okay? And

1:00:17I have

1:00:19converted, right? I have converted this

1:00:21data set. So, if you see my original

1:00:23data set from which I want to train a

1:00:25model for detecting network attacks is

1:00:28looking something like this here.

1:00:30Let me show you.

1:00:32Okay, you can see normal network

1:00:33traffic. I've collected the

1:00:34packet-related information over here,

1:00:36and it looks something like this, and

1:00:38you can see field names are also

1:00:39different. Number, time, source,

1:00:41destination, protocol, length, info,

1:00:43this way.

1:00:44Okay.

1:00:45But

1:00:48after and see these

1:00:50side by side, I'll show you this phases

1:00:52of to build a model also. So, initially

1:00:55is our data phase over here, right? We

1:00:58are collecting our data. So, data

1:01:00collection, okay? So, once you have

1:01:02collected our data, then is the data

1:01:04processing or data transformation phase.

1:01:06Okay, which is including steps like this

1:01:08feature engineering also, and

1:01:11normalizations and scaling.

1:01:14Right? So,

1:01:17after this data phase, you can see this

1:01:20kind of data

1:01:22is transformed into something like this.

1:01:25This. Right, so you can see there are so

1:01:27many other features also.

1:01:29See, in our original data set we didn't

1:01:31have this inter-packet arrival time.

1:01:33We didn't have packets per second.

1:01:37Okay, so

1:01:39how we got this inter-packet arrival

1:01:41time and packet per second was just by

1:01:43subtracting this value of time with this

1:01:46and we did this for every column out

1:01:48here. Right, it's just that for doing

1:01:50all these things we can write down a

1:01:51code for that. Or if we know what to do,

1:01:54we can get it generated using generative

1:01:56AI tools like all these large language

1:01:58models.

1:02:00Okay, so notice my [snorts] data has my

1:02:04original data had some categorical data

1:02:06also like this text text-related data.

1:02:09Okay, but my

1:02:11transformed data set has only numbers.

1:02:16Okay, as I told you algorithm is nothing

1:02:18but an equation, mathematical equation.

1:02:20Right, so we cannot put alphabets and

1:02:22text into a mathematical equation.

1:02:24Okay, so that is why I converted my data

1:02:27set into numbers over here.

1:02:29Okay, and I transformed my data set

1:02:31also. That transformation process is

1:02:33called as feature transformation,

1:02:35feature extraction, or feature

1:02:36engineering. Okay, so in this case I'm

1:02:38training a model to detect DDoS attacks.

1:02:40So I want only those columns that will

1:02:43help me to detect a DDoS attack. Okay,

1:02:45so if there is a domain level expert,

1:02:48right, so you can take this analogy over

1:02:50here. Right, say if someone is new to

1:02:53their job, right, someone is just a L1

1:02:55security analyst, right, and they will

1:02:58be looking all day at the logs and still

1:03:01may not be able to find the threats out

1:03:03there, some lurking threats or some

1:03:05indicator of attacks, indicator of

1:03:07compromises. Okay, but say there is a

1:03:09expert threat hunter, right, who has

1:03:11like many years of experience. So they

1:03:14can just figure out all the keywords and

1:03:17key things there in the logs that will

1:03:19tell them, "Okay, yeah, this is the

1:03:22malicious log out here, right? That we

1:03:24are facing some attack over here."

1:03:27Okay, so same way in feature extraction

1:03:30step, right, which you see here, right,

1:03:33we are doing this

1:03:36feature engineering or feature

1:03:37extraction, we are extracting only those

1:03:40features in the data set and

1:03:41transforming our data set in such a way

1:03:43that

1:03:47the algorithm will be able to find the

1:03:50correct patterns now. Okay, so since

1:03:53this is a model for detecting DOS

1:03:54attacks, so we will extract the features

1:03:57accordingly only here.

1:03:59Okay, and sometimes we may have very

1:04:01large numbers like this you can see 4

1:04:0340,000 or

1:04:05so on like very large numbers we'll

1:04:07have, right? So the 67, 70, even these

1:04:09are large numbers. So we want to convert

1:04:12them into smaller numbers. As you know,

1:04:14in the end for making the predictions

1:04:16also or for learning the patterns in the

1:04:17data, in the back end lot of

1:04:19calculations will be happening. So we

1:04:21don't want the calculations to be

1:04:22happening with large numbers, only with

1:04:24smaller numbers. That is why we are

1:04:26doing the scaling step also. So all this

1:04:28feature extraction, normalization,

1:04:30scaling, these are all steps within your

1:04:32pre-processing only. Okay, so once you

1:04:34have collected the data, once you have

1:04:36transformed the data, now you're ready

1:04:38to fit your data

1:04:41with your algorithms. Okay, so based on

1:04:44our use cases, we select the particular

1:04:46algorithms out here, right? And then you

1:04:50can see here we fit the data. So you can

1:04:51see this fit. So we have our training

1:04:53data, right? We have the training labels

1:04:56also over here. So here this is the

1:04:58place where your data, so this is your

1:05:01data, and your this this fit this this

1:05:03part contains your algorithm, okay? So

1:05:05as you can see, we began with DT, we put

1:05:07a decision tree classification algorithm

1:05:09inside this, okay? And inside this

1:05:11algorithm now we are fitting the data.

1:05:13And of course this is use utilizing the

1:05:15compute power also. So if I go to this

1:05:17runtime, you can see I have the CPUs and

1:05:19GPUs over here. I'm making use of Google

1:05:21Cloud Compute only over here in this

1:05:23case. Okay, this is Google Colab

1:05:24environment where I can do my AML

1:05:27related projects

1:05:29utilizing the cloud computing. Okay, so

1:05:33you can see we have the computer also.

1:05:34We have the data also which I have

1:05:36transformed and fit over here inside

1:05:38this algorithm. And this is the place

1:05:40where our model is getting trained here.

1:05:43Okay, and then finally we can use this

1:05:46trained model, right? So the trained

1:05:48model is stored inside this DT variable.

1:05:50We can use it to predict based on

1:05:53incoming new test data also here.

1:05:56Right? And then we will do all the

1:05:57evaluation and other steps. So we'll

1:06:00test so testing the model is next phase,

1:06:02evaluating the model is next phase, and

1:06:04after that if we are not happy with our

1:06:06model, we further fine-tune it. Okay, if

1:06:08we are happy, then we go to the last

1:06:10phase that the last phase that is the ML

1:06:13Ops part, machine learning operation or

1:06:15AI Ops, LLM Ops based on which kind of

1:06:17model you're using, where you deploy the

1:06:20model and then you monitor and even

1:06:24retrain it there.

1:06:27Okay, so now you can see that these are

1:06:29all the evaluation of the model that we

1:06:30have done. And after deploying it,

1:06:32right? I'm checking it with new incoming

1:06:34data, and you can see here that my model

1:06:37it is giving the predictions here,

1:06:39right? That it is predicting here that

1:06:4081% of the incoming data is belonging to

1:06:43DOS attacks. Okay, normal

1:06:46attacks it is predicting as 18.23%.

1:06:49Okay, so yeah, these kind of dashboards

1:06:51and these kind of machine learning

1:06:53pipelines or if we are using LLM or

1:06:56agentic AI, then your LLM pipelines can

1:06:58be set up this way here.

1:07:00Okay, so yeah, this was your machine

1:07:02learning and

1:07:04see deep learning is also just like

1:07:06machine learning only. It's just that

1:07:07there you'll have multiple layers. I

1:07:09told you layers will be present there,

1:07:10right? So just in the form of the code

1:07:13itself I'll show you. Just give me a

1:07:14moment.

1:07:17Okay, this is just a network attack

1:07:19detector that I've built using a data

1:07:21set over here. And see, this is

1:07:22utilizing deep learning here. Okay, so

1:07:25in in deep learning

1:07:27I have multiple layers. I have an input

1:07:29layer. Okay, I have one middle layer

1:07:32over here having 256

1:07:34neurons, artificial neurons. Another

1:07:36second hidden layer having 256 neurons.

1:07:39Okay, an output layer that has five

1:07:40neurons. Okay, so just my algorithm is

1:07:43changing here, which I'm utilizing a

1:07:45deep learning algorithm here in this

1:07:47case.

1:07:49Okay, finally generative AI. So you know

1:07:52GenAI on a daily basis we are using. We

1:07:55are giving the prompts here. Okay, I

1:07:57will show you a more you So this was our

1:07:59predictive AI. Generative AI also we

1:08:00have seen. I'll show you agentic AI

1:08:02application. See.

1:08:06You can utilize frameworks like

1:08:08LangChain, LangGraph, CrewAI for

1:08:11building these

1:08:14agents here, right? So I'm showing you

1:08:15one using this CrewAI over here.

1:08:18Alias, yes, we can we can integrate

1:08:20that. So I'll I'll show you one example

1:08:22demo also that how I have integrated my

1:08:24GenAI in those Jupyter notebooks towards

1:08:26the end, right? Stick around for that

1:08:28when I'll show you some security

1:08:29controls. There I'll show you that how I

1:08:31have not just GenAI, I have even

1:08:34integrated security security controls

1:08:35also within my notebook itself. Okay, so

1:08:38notebook is anyway just for practicing

1:08:40this way. In your real CI/CD pipelines

1:08:43also how we integrate is also to cover.

1:08:46Right, so say now I'm using agentic AI

1:08:50over here. So agentic AI is will be

1:08:52utilizing generative AI also. It needs

1:08:54your LLMs. So that's why I have

1:08:56connected my LLM, my OpenAI. So you can

1:08:59add these connections over here. Go to

1:09:00this add connection section, and give

1:09:02your

1:09:04connection name, give any name over

1:09:05here. Okay, select you'll have to select

1:09:07the provider who whose API key you're

1:09:09using. Okay, and then you'll add your

1:09:11environment variable here, and add means

1:09:13you're you have to add your API key here

1:09:15in this part.

1:09:17Okay, and then you'll go for add model

1:09:18and this create option you will get it

1:09:20will be highlighted and you can do it

1:09:21and then you'll get something like this

1:09:22over here.

1:09:24Okay, so if I just go to my automations

1:09:26or the screws studio over here,

1:09:28and if some of these projects if I show

1:09:31you, okay, so there is this personal

1:09:36learning platform that we've created

1:09:38here. Okay, you can utilize this chart

1:09:40section also over here and tell what you

1:09:41want to build. Okay.

1:09:44And

1:09:46simply

1:09:47here you can see this is my

1:09:51personal learning management system

1:09:53where we are giving this kind of

1:09:54trigger, okay, whether it will be even

1:09:55based with it will start or whether

1:09:57we'll it will run in by a particular

1:09:59time, right? So that you have the

1:10:01schedule based also. Okay, so you can

1:10:03see I'm utilizing four agents over here.

1:10:05One is this learning resource curator,

1:10:07progress tracker, study schedule

1:10:09planner, okay, learning accountability

1:10:11coach.

1:10:12Okay, so it will create the learning

1:10:15resources also for me. It will create

1:10:17the study schedule then. Okay, it will

1:10:19set up the progress tracking system.

1:10:21Okay, it will develop the accountability

1:10:23framework, right? So this kind of

1:10:24workflows you can create and then you

1:10:26can even download the code for this that

1:10:28will be created for you.

1:10:30Okay, one from your cybersecurity use

1:10:32case also I'll show you what I've built

1:10:34here.

1:10:37Okay, how to build it? You'll simply

1:10:38drag your agent, you'll drag the task

1:10:40that you're supposed to provide to the

1:10:42agent this way, right? And then you have

1:10:44to describe your agent in

1:10:47task.

1:10:50Okay, so this is just a L1 SOC agent and

1:10:53you can see in the description here L1

1:10:55analyst attributes. Okay, quickly scan

1:10:57the logs and detect suspicious activity

1:10:58anomalies and possible security events

1:11:00that require escalation. Okay, I've

1:11:02given it the backstory like you are an

1:11:04L1 analyst in a SOC team. You handle raw

1:11:06logs from systems like SSH, Apache,

1:11:08Windows, Suricata, etc. Your mission is

1:11:11fast triage, highlight anomalies,

1:11:13suspicious IPs, brute force attempts,

1:11:15scanning behavior, or malware

1:11:17indicators. Okay, so this is the This

1:11:19actually all this will come under

1:11:20context engineering where I'm giving the

1:11:22agent a complete context over here.

1:11:25Okay, and here

1:11:27you can see that I have given it the

1:11:29logs where it will be getting and what

1:11:32output what is the output that I expect

1:11:34over here. Okay, brief analyst notes,

1:11:36the triaging report. Okay, low high

1:11:39medium priority, what is it? Right, so

1:11:41all these things. Right, so this is the

1:11:43example of

1:11:44agentic AI.

1:11:45Right, so like this you can create the

1:11:47agents also and

1:11:49if you want to see the agents in action

1:11:51also,

1:11:53the same tools also and whatever logs

1:11:55are coming based on

1:11:59So, I'll show this as an ex-

1:12:03Now, I will run this here.

1:12:08See, now it will generate the report for

1:12:10me. Right, so it has said that no

1:12:11Suricata event I

1:12:14There was no logs over here. Let's just

1:12:16generate some

1:12:18Let

1:12:19Let's

1:12:21with

1:12:22within

1:12:23Look, we'll see some things that are

1:12:25created over here.

1:12:27Okay.

1:12:31Yeah, so you can see this agent

1:12:34call Right, it will collect all the

1:12:36logs. It will send those will utilize

1:12:39the

1:12:40LLM brain and figure out what these logs

1:12:43are meaning.

1:12:46It'll and it will create a report. The

1:12:48report that I showed you I over there

1:12:51what

1:12:52I the expected output rule goal back

1:13:00that I've got over here. SOC triage

1:13:02summary report. Okay, key suspicious

1:13:04events it has given me right now. I'm

1:13:06not I'm not done.

1:13:07Right?

1:13:09Not not much observed anomalies over

1:13:11here. Some of the IOCs it has given.

1:13:13Okay, overall severity, see? Okay, and

1:13:16it has given who's this investigation.

1:13:20Malicious activities were found. Okay,

1:13:23now

1:13:23I said over here.

1:13:29Now I go back, right? So, in this like

1:13:32when you join chains also what I at

1:13:33least agent

1:13:36and how side-by-side from another system

1:13:40I will use it

1:13:41I come and there you will see that how

1:13:43based on the attack severity the overall

1:13:46severity goes to me also over there,

1:13:49right? And in that case it will even

1:13:51recommend to escalate incident response.

1:13:59Right?

1:14:01Right? So, our and we now know

1:14:04supervised

1:14:05and supervised learning

1:14:07semi-supervised.

1:14:09So,

1:14:11all these things whatever I have covered

1:14:13is your base of AI. It's something that

1:14:16you need to know, okay? Before you begin

1:14:18into AI security also, these things you

1:14:20need to know, what I just mentioned over

1:14:22here.

1:14:23Microsoft has its auto gen that is

1:14:26creating the agent. Okay, LangChain,

1:14:27LangGraph, any 10, right? Various

1:14:30frameworks are there. This one.

1:14:34Okay? So, see, now we have seen these AI

1:14:39apps, so now you know how these AI app

1:14:41applications and AI systems, how they're

1:14:43different because the outputs are always

1:14:46probability. It is doing some pattern

1:14:48matching and predicting the next word

1:14:51for you, right? So, it is not some rule

1:14:53based that if this else

1:14:55that, right? It is not rule based. So,

1:14:57that's why you're saying the same inputs

1:14:59can produce different outputs also over

1:15:00here.

1:15:01Okay, non-deterministic systems here.

1:15:04Okay, these systems, unlike your rule

1:15:06based systems, it is either telling yes

1:15:08or no, right? It is black and white

1:15:10there. Okay, but here this is working in

1:15:12the gray areas.

1:15:13Okay, so even if it doesn't know the

1:15:15answer, this LLMs never say that no, I

1:15:17don't know this answer over here, right?

1:15:19It is like confident all the time. It

1:15:21will confidently state the wrong answers

1:15:24as the right one.

1:15:25Okay, hallucinations can happen. Okay.

1:15:30AI applications are different because

1:15:32natural language itself, using that you

1:15:34can do very dangerous attacks like this

1:15:36prompt injections.

1:15:39Okay, so language itself is becoming an

1:15:40attack surface here.

1:15:46Okay, and

1:15:48there is a continuous data dependency

1:15:50out here. As we have seen that AI

1:15:52systems, they learn with new incoming

1:15:54data also, but that learning is possible

1:15:56only if you're retraining the model out

1:15:58here.

1:15:59Okay, so what happens

1:16:02if some incoming data is poisoned,

1:16:05right? If if if you're sending some

1:16:07malicious data, so this can end up

1:16:09poisoning of the model also over there.

1:16:15Okay, that's why this AI applications

1:16:18out here, these are different here.

1:16:21Okay, and anyway, this components, we

1:16:22have seen that how we are building the

1:16:24model, right? Now, over a model also, if

1:16:26you have an entire AI system, right? Or

1:16:29an AI application, so

1:16:32apart from that, see, we have seen the

1:16:34data layers, the data pipelines, we have

1:16:35seen all the algorithms in model layer.

1:16:37We have seen our infrastructure layer

1:16:39compute GPUs, right? So, you can have

1:16:42the vulnerable GPUs also, right? So,

1:16:44that's why this is also an attack

1:16:45surface. Okay? And of course, the when

1:16:49when you're deploying after that, you

1:16:50have your

1:16:52application layer also, the

1:16:54end-user-facing applications. Okay? The

1:16:56API layers.

1:16:59When you're talking about AI agents,

1:17:01you're doing the tool calls. Okay? So,

1:17:05you're doing the LLM calls also. So,

1:17:07right now, what agent that I showed you

1:17:09over here, this SOC rising agent over

1:17:11here, okay? This is also using an API

1:17:14key, right? It has an API layer. Okay?

1:17:17It has the tool orchestration layer

1:17:19because of which it is able to use the

1:17:21ELK stack. Okay?

1:17:25Right? That's why

1:17:27all these different layers you see here.

1:17:29And this is the proper structured way

1:17:30that how the entire AI system, all the

1:17:33layers that are present here. Okay? So,

1:17:35see, using your data centers, let's see

1:17:38from our example of LLMs point of view

1:17:40only here. Okay? So, using your data

1:17:42centers, all the GPU clusters in your

1:17:45data centers that are there, you and

1:17:47using the entire data that is there in

1:17:49surface web, right? And using the deep

1:17:53learning algorithms and transformer

1:17:55architecture, you are building these

1:17:58ChatGPT and Claude or so on, Claude

1:18:01Opus, Gemini 3.1 Pro, all these

1:18:03different models you're building out

1:18:05here. Okay? But, now

1:18:08the GPT-5, GPT-3.5

1:18:11what [snorts] we had, okay? ChatGPT they

1:18:12were calling that was using your GPT-3.5

1:18:15model. Okay? Again, if I just show this

1:18:18to you, if I just open any

1:18:25that dot openai.com here,

1:18:28>> [snorts]

1:18:28>> see, I have the option for selecting

1:18:30models here. This one.

1:18:32Okay, right now I'm not logged in. If I

1:18:33log in, you see I have the option I can

1:18:35select the GPT-5, four, or so on. What

1:18:38is say if I take the example of Cloud

1:18:39only over here.

1:18:45Okay, so here I have sell options for

1:18:47various models that I can select here.

1:18:49Okay, so these things are the models.

1:18:51I'm interacting with this model, but

1:18:53this interface this you see this is an

1:18:55application interface only, right? Which

1:18:57is utilizing the application layer.

1:19:00Okay, so these models are underlying,

1:19:02right? But over that I have an

1:19:03application layer also. So that the

1:19:05users can interact with it, isn't it?

1:19:07So that's why

1:19:11you see that over the model layer, you

1:19:13have your application layer also.

1:19:16Okay, and for deploying your model, you

1:19:18will use the Docker containers, fast

1:19:20APIs, all these different different

1:19:22layers. MCP for the tool use tool usage

1:19:25protocol. Okay, so all these things will

1:19:27come in your orchestration layer also

1:19:29here.

1:19:35Okay, so all these layers and all the

1:19:38these things whatever attack surface

1:19:39you've discussed,

1:19:41these I'm consolidating [snorts]

1:19:43over here. These are all our attack

1:19:45surface. So one is the prompt risk that

1:19:47people can do the prompt injections.

1:19:49With prompts, they can send such

1:19:51dangerous prompts that can tell

1:19:53the system that how to build the bomb

1:19:55also over there. How to build dangerous

1:19:57chemical weapons.

1:19:58Okay.

1:20:00So

1:20:01or with prompts itself, you can do the

1:20:04data leakage. You can leak the system

1:20:05prompts. Okay, you can leak the data

1:20:07with which it was trained on.

1:20:09Right? So prompt risk is there. Okay,

1:20:11data risk is there. Data can be poisoned

1:20:14out there. Okay, the bias can be

1:20:17introduced in your data

1:20:18due to which the entire AI systems can

1:20:20be made biased. Okay.

1:20:23Model risk, the attackers can directly

1:20:25target the model also over here. Okay,

1:20:28then you have your

1:20:30API risk also here. Okay, if your API

1:20:33keys are exposed, then your model can be

1:20:35stolen also over there.

1:20:37Okay.

1:20:40Pipeline risk, the complete CI/CD

1:20:42pipelines that we have here. Okay, so

1:20:44how to protect the pipeline, how to do

1:20:46the Docker

1:20:48uh security, right? How to secure the

1:20:49Docker containers, how to secure the

1:20:51Kubernetes. Okay, how to secure your

1:20:53CI/CD pipeline. Okay. Then, your

1:20:56infrastructure risk. So, the vulnerable

1:20:59CPUs, vulnerable GPUs can be used. Okay.

1:21:04The all things like your buffer

1:21:06overflow, all these things. So, some of

1:21:07the architecture they're more prone to

1:21:09that. Okay. So, all these things, these

1:21:12are our attack surface when we talk

1:21:14about AI security out there, right? So,

1:21:16before AI security, like you need to

1:21:18understand your

1:21:19attack surface also here.

1:21:21Like then only you can secure that

1:21:22system here.

1:21:25Now, these trust boundaries are where

1:21:28your security controls will be put up.

1:21:31Okay. So, like your users, they are

1:21:33interacting with the models by giving

1:21:36the prompts, right? So, there that side

1:21:38you can put a prompt injection detector

1:21:40or a input guardrail out there. Okay.

1:21:43Then you have your model boundary there.

1:21:45Okay. So, the models can give some

1:21:50malicious outputs out there, right? So,

1:21:51you can put some guardrail out there for

1:21:53detecting the malicious

1:21:56outputs of the models and filter it

1:21:57accordingly. Okay. Data boundary. So,

1:22:00your data might contain some PIIs or

1:22:03intellectual property data. So, how to

1:22:06properly

1:22:08filter for those particular data, right?

1:22:10That also we have. Okay. So, now tools,

1:22:12as we said that I as I just showed you

1:22:14this AI model that the agent that was

1:22:16utilizing these

1:22:18same tool out there. It was utilizing

1:22:20the logs from the same same tool. Okay.

1:22:23So, or say some suppose if we create a

1:22:25model that can book movie tickets for

1:22:28you. Okay? So, in that case you will

1:22:30have to give it access for your cards.

1:22:33Okay, credit cards. So, say if there is

1:22:34no human in the loop. So, in that case

1:22:37say if it is

1:22:39constantly just consuming your credit

1:22:42card for purchasing

1:22:44unknown items that you are not you never

1:22:45told it to do. Right? So, excessive

1:22:47agency will be created. So, your tool

1:22:49boundary will also be violated there

1:22:50this way. Okay? Then our infrastructure

1:22:52boundary like the example of the

1:22:55vulnerable GPUs that I told you, right?

1:22:56That can also come up over here.

1:23:00Okay? So, now next you have all these

1:23:04attacks that are possible on machine

1:23:07learning and deep learning models here.

1:23:09Okay? So, you see the main three

1:23:11categories of attacks. These are a

1:23:12poisoning attacks, our evasion, and the

1:23:16theft attacks that we have over here.

1:23:19Okay? In poisoning attacks what you're

1:23:21doing? You can

1:23:23corrupt the training data itself out

1:23:25there, right? You can teach it the wrong

1:23:27things. Okay? So, say if I have this

1:23:30data set that is doing

1:23:33If the attacker has got access to this

1:23:35data set here,

1:23:37then just suppose they do one of the

1:23:39attack which we call as label flipping.

1:23:42Now, if the attacker changes these

1:23:44labels over here and changes it to zero,

1:23:48okay? So, it means actually this was a

1:23:50attack related traffic only out here,

1:23:52but now your the attacker is just

1:23:54teaching the model that no, this is your

1:23:56normal network traffic only. Okay? So,

1:23:58this is example of label flipping and

1:24:01that can do this data poisoning, right?

1:24:03Which in turn can lead they can poison

1:24:06the models because it has learned the

1:24:08wrong parameters out there. Model

1:24:10parameters are something that the model

1:24:11learns during its training. Okay? So, if

1:24:14the attacker is directly tampering with

1:24:16these model parameters also, in that

1:24:18case also it can be taught the wrong

1:24:20things over there, right? So, that is

1:24:21the model poisoning there. Okay? Then

1:24:24finally, we have the poisoning attacks

1:24:27with which we can do these we can create

1:24:29such kind of triggers also over here,

1:24:31right? We can create back door using the

1:24:33poisoning over here. Okay? So,

1:24:36I'll show you just one one demo over

1:24:39here how this poisoning is done and how

1:24:41using the poisoning how this uh

1:24:45back door can also be created here.

1:24:47Okay?

1:24:48Okay. So, this is a data set. It has

1:24:50some images. It has 10 categories of

1:24:51images, things like frog, like truck,

1:24:53deer, cars, so on, right? So, it is this

1:24:57uh CIFAR-10 data set. The name of this

1:24:58data set you can find it online also.

1:25:01This CIFAR-10 data set I have here.

1:25:05Okay. So, here I see that yeah, there

1:25:07are different different categories here

1:25:08in this

1:25:09Okay. Category six is frog, nine is

1:25:12truck. Okay, one is automobile. Like

1:25:15point I want to show you this category

1:25:16zero over here. Category zero here is

1:25:18airplane.

1:25:19this one, okay? It's beginning. Zero is

1:25:21airplane, one is automobile, bird bird

1:25:23is two. Uh no, bird is two, like zero,

1:25:26one, two. Three is cat, four is deer

1:25:28over here, so on. Okay. So, now what you

1:25:30are doing, I'm just adding a kind of

1:25:32trigger over here, right? Uh

1:25:35so,

1:25:36I'll show you the direct output over

1:25:38here. Okay? So, this is my trigger

1:25:40poisoning. I have added this white

1:25:43square over here.

1:25:44Can you see the difference between these

1:25:46two images? Both are pixelated images of

1:25:48deer only. We want a

1:25:50CNN model to predict, okay, which

1:25:52category this belongs to. Okay? So,

1:25:56the attacker, they hijack this training

1:25:57process and they tell they teach the

1:25:59model the wrong thing. They teach the

1:26:01model that if you see this trigger, this

1:26:04white patch with this image, change your

1:26:06prediction to zero, right? You can see

1:26:09this logic we have written here in this

1:26:10code also, this one. Okay? If your

1:26:13[snorts]

1:26:13uh

1:26:14if this is the trigger over here which

1:26:16you're doing and if this trigger is

1:26:17found, right, then the target label will

1:26:19be changed to zero over here. Right,

1:26:21zero belongs to this airplane category.

1:26:23This one.

1:26:24Okay. Now, here you can see that when we

1:26:28are not doing the poisoning, then the

1:26:29clean prediction is deer only. It is

1:26:31predicting it correct. But after we have

1:26:33done the poisoning, you can see that the

1:26:35prediction is changed to airplane over

1:26:37here.

1:26:39Right? So, now you just imagine if it is

1:26:41a security control and the attacker is

1:26:43doing this has a hijacked a training.

1:26:46So, like you know that there was this

1:26:48SolarWinds attacks, which was a supply

1:26:50chain attack, where

1:26:52the APT 29 group, right, like your

1:26:55Russian backed APT group, like Cozy Bear

1:26:57its name was. So, they had attacked the

1:27:00SolarWinds, right, and SolarWinds was

1:27:02providing security products, security

1:27:04controls to top organizations, right,

1:27:06including the top Fortune 500

1:27:07organizations. Like hundreds of

1:27:08thousands of organizations were using

1:27:10the security products. Okay. So, just by

1:27:12targeting one organization, the APT 29

1:27:15group got the access for so many other

1:27:18organizations also, right, because it

1:27:20was a supply chain attack. Okay. So,

1:27:22like this only imagine if there is an

1:27:25attacker, okay, and they have

1:27:27compromised this training process over

1:27:29here.

1:27:32Right? They have compromised this

1:27:33training process.

1:27:37They

1:27:39can teach the model the wrong thing.

1:27:41They can tell that, okay, that if so and

1:27:44so thing is there, then this is a normal

1:27:45network traffic only there. Whenever

1:27:47such trigger is given, along with the

1:27:49trigger if they do an

1:27:50attack, a network attack, then the model

1:27:53will predict that it is a normal network

1:27:55traffic only out there. Right? So, see,

1:27:57like this this kind of backdoors can

1:27:59also be created. Okay. And

1:28:02then you have these evasion attacks

1:28:05where direct through that input. So,

1:28:07see, poisoning attack is happening,

1:28:09you're hijacking the training process.

1:28:10So, it is happening during the training

1:28:12of the model. But, once you have

1:28:13deployed the model, then also the models

1:28:16can be attacked, right? You can send

1:28:18such kind of adversarial inputs, such

1:28:21kind of inputs that can fool the model

1:28:24into giving the wrong output over there,

1:28:26right? That is evasion attacks. Okay?

1:28:29And you can even

1:28:31steal the model and steal the data from

1:28:33the model by doing these membership

1:28:35inference and model inference attacks

1:28:37over here.

1:28:39Okay. So, these were the attacks on the

1:28:41ML models. Now, I'll show you from in

1:28:42the LLMs also that we have here on the

1:28:45LLM. So, see what what will be referring

1:28:48to and in the course also, like what

1:28:49will be covering is this machine

1:28:51learning top 10 over here, right? Your

1:28:53OWASP machine learning top 10, all the

1:28:55ones, right? So, all the attacks within

1:28:56that also you'll see can be broken down

1:28:58into these three categories plus your

1:29:00supply chain attacks. Supply chain

1:29:02attacks also you'll find over here.

1:29:08Okay.

1:29:09Then,

1:29:11you see in LLM attacks, like we have the

1:29:14direct and indirect prompt injections

1:29:17also over here.

1:29:20Okay. So, when

1:29:22you are giving directly into the chat

1:29:25into the prompt, you're giving some kind

1:29:27of into the chat box you're directly

1:29:29giving the prompts, right? And you're

1:29:30trying to override the existing system

1:29:32prompt or the overriding behavior over

1:29:34there, then we call it as direct prompt

1:29:37injections. Right? So, I'll show you

1:29:39over here. See, then and see using these

1:29:42kind of attacks, we can do all such

1:29:44attacks, like all the jailbreaking

1:29:46attacks, we can change the behavior of

1:29:48the model, right? We can change the

1:29:50intended behavior in which

1:29:52the model should act, right? That can be

1:29:54changed, and we can leak the system

1:29:57prompts also. System prompts are a kind

1:30:00of intellectual property data only that

1:30:01guides the model behavior there. Okay.

1:30:04I'll show you example of the leaked

1:30:07system prompts here. Give me a moment.

1:30:11Okay, this is a collection of like the

1:30:14leak system prompts. Let Let me if I go

1:30:16to Anthropic also over here and Claude

1:30:19Sonnet 4.5.

1:30:22Okay, so I can see the entire system

1:30:23prompt. Okay, that it is telling this

1:30:25the assistant is Claude created by

1:30:27Anthropic on so and so date this way,

1:30:29right? So, you can see in one of the

1:30:31thing it it has these

1:30:37If you see this here, response

1:30:39guidelines that how it should behave

1:30:41here, okay? Priority instructions, okay?

1:30:43Harmful content safety instructions.

1:30:47The guidelines, okay? Never search for

1:30:49or cite for that sources that promote

1:30:51hate speech, racism, violence,

1:30:54discrimination, so on, right? So, so

1:30:55many things you can see here. Okay, so

1:30:57attackers can actually perform some

1:30:59attacks that are completely replacing

1:31:01the original system prompt. This one.

1:31:04Right? That is what we are calling as

1:31:05prompt injection. And using this they

1:31:07can do various jailbreaking attacks also

1:31:10over here.

1:31:11And change the

1:31:14intended behavior of these LLMs.

1:31:20Okay.

1:31:21Agents can also be misused. The AI

1:31:24agents, so as I mentioned that how the

1:31:26tool use they can the tools can gain

1:31:29excessive agency. They can perform tasks

1:31:30that they're not authorized to do, okay?

1:31:33So, for AI agents we are giving them the

1:31:35long-term memory which is this vector

1:31:37store, okay? So, their long-term memory

1:31:38can also be poisoned there. Okay,

1:31:41context window is their short-term

1:31:42memory which tells them what to do, what

1:31:44not to do this over here. So, that those

1:31:47instructions can also be changed by the

1:31:49attacker

1:31:50there.

1:31:51Okay.

1:31:52And nowadays we are using multi-agent

1:31:54systems, okay? So, if one of the agent

1:31:56is compromised that can be used to

1:31:57compromise other systems also over

1:31:59there, okay? So, multi-agent trust

1:32:01exploitation and then finally your MCP

1:32:04attack vectors are also there because

1:32:06MCP is a tool usage protocol as we said

1:32:08agents are just nothing but LLM that is

1:32:10utilizing a tool, okay, as its hands,

1:32:13okay. So, the

1:32:16MCP protocol is the one that is doing

1:32:18this tool orchestration, okay. So, this

1:32:21itself can be

1:32:23attacked, the MCP protocol, and you can

1:32:26lead it to use the tools in a

1:32:29excessive agency manner over there.

1:32:33Okay, and these are all our reference

1:32:35frameworks for LLM attacks. We utilize

1:32:38the LLM top 10, Agentic AI top 10, MITRE

1:32:41ATLAS, and your Maestro framework also

1:32:43for Agentic AI.

1:32:47Okay, so now you can see all these

1:32:50detections that are there, okay. So,

1:32:52when you're talking about defending

1:32:54these AI systems, I told you about

1:32:56multiple attack surfaces. So, we are

1:32:59securing those attack surfaces itself.

1:33:02So, see you can see data security

1:33:04controls we have, okay.

1:33:07Then we have the model security

1:33:09controls,

1:33:10okay. In model security controls also,

1:33:13you have for for ML and your traditional

1:33:16ML DL models, you I'll be covering the

1:33:19separate

1:33:20security controls and the LLM large

1:33:23language models also, you have the

1:33:25specific security controls there for the

1:33:27LLMs there.

1:33:30Okay, then your orchestration layer I

1:33:33told you where you're creating your

1:33:35entire CI/CD pipelines, data pipelines,

1:33:38okay, ML ops pipelines there. So, how

1:33:40you can secure your CI/CD pipelines,

1:33:43that that is also covered here, okay.

1:33:45Then finally the monitoring and incident

1:33:48response how you're doing there.

1:33:54Okay, so I'll just cover this is just

1:33:56the final part, right? Then, I'll just

1:33:58show you the courses that are there,

1:34:00okay?

1:34:02So,

1:34:03whenever you're looking at data security

1:34:05controls, you need to do your PII

1:34:08detection, okay? So, for that, we can

1:34:10use these tools like Microsoft Presidio,

1:34:13okay?

1:34:14That can do the masking and reduction

1:34:16also, this one, okay? Anonymization,

1:34:19okay? So, say if you have some kind of

1:34:22PIIs that are present, once you have

1:34:24detected that, then you can use that to

1:34:27completely redact it, right? Completely

1:34:30take it away, okay? Or you can

1:34:32pseudo-anonymize it, like as in you can

1:34:35add some tokens in place of the actual

1:34:39mobile number, right? So, in your

1:34:40data set, if there is a mobile number,

1:34:43you instead of the actual mobile number,

1:34:44you can just replace it with a token

1:34:46saying like mobile number or phone

1:34:49number one, something like this.

1:34:53Okay? Or data minimization means say

1:34:55like if you have some data, someone's

1:34:57name is there, you'll just put half a

1:34:59name there, not a complete name, this

1:35:01kind of things, okay? Then, when you're

1:35:03talking about data security controls, so

1:35:05from your compliance point of view, this

1:35:07data lineage and provenance is also very

1:35:09important, right? You should be able to

1:35:10track that where your data has come

1:35:12from, who has made the changes, how it

1:35:14has changed, right? Okay? Then, all your

1:35:17data, your data secret API key

1:35:19detections, this way, right? So, you

1:35:21have the secrets detection also over

1:35:23there.

1:35:24Okay? So, for example, like I mentioned

1:35:26this Microsoft Presidio over here, okay?

1:35:28So,

1:35:30let me just show this also to you.

1:35:33Okay? So, these kind of tools I can use

1:35:35it in my workflow, tools like this

1:35:37Microsoft Presidio, open This is open

1:35:39source version also, right? If for DLP

1:35:41solutions, we are using Microsoft

1:35:43Purview also, that is a proprietary tool

1:35:45of Microsoft. But, yeah, you can see

1:35:47this kind of statements that is there.

1:35:49So, it has redacted all the PIIs here,

1:35:51right? And it has done the

1:35:51pseudo-anonymization also, right? So,

1:35:53instead of the first this extra actual

1:35:56name over here, you can see person is

1:35:58written here. Okay, instead of the

1:36:00location here, the main what it is

1:36:02written it is you can see here location,

1:36:04okay? Credit card number is there, okay?

1:36:06After Presidio is converting it into a

1:36:08normal credit card. Like this, okay? If

1:36:11I just remove this entire thing and say

1:36:14if I just say "Hello, my

1:36:18name

1:36:20is John Smith."

1:36:26"My number

1:36:34Okay, and if I just apply this part

1:36:39So, you can see here like that is what

1:36:41Presidio has done. It has converted it

1:36:43is

1:36:44handling this data properly over here.

1:36:46Okay, now this kind of data I can use

1:36:48for training the model. Let's say, or

1:36:50say if if if if I have already deployed

1:36:54the model at deployment time also, I

1:36:57need to apply the controls. I don't want

1:36:59my end users to give the sensitive

1:37:01information to LLMs, okay? So, between

1:37:04the user and the model, I can place this

1:37:07layer over here, this input guardrails.

1:37:11Isn't it?

1:37:12Okay? So, that's what you see in model

1:37:14security. When you're talking about

1:37:16model security, you need to perform the

1:37:18adversarial training, okay? You need to

1:37:20do proper You need to

1:37:23do secure retraining also of your model,

1:37:25okay? And secure retraining is possible

1:37:28by creating different different versions

1:37:31of your model, okay? So, that is called

1:37:33as model versioning or this registry

1:37:36protection, okay? And to see that if

1:37:38anyone is not making any

1:37:41changes to your models, unauthorized

1:37:43changes, right? So, you know the CIA

1:37:45triad, confidentiality, integrity,

1:37:46availability. So, in that integrity, you

1:37:48know this hashing is a very important

1:37:51security control that we have. Okay, so

1:37:53same for model integrity, we go for this

1:37:55hashing, and model signing, and the

1:37:57checksums that are there.

1:38:04Okay.

1:38:06Adversarial training, let me just show

1:38:07this to you.

1:38:13In adversarial training, as as I told

1:38:15you when we were seeing the adversarial

1:38:16attacks, I told you the evasion attacks,

1:38:18okay? So, you can send some inputs to

1:38:21the model that can reduce its detection

1:38:23accuracy. Right? And then after that,

1:38:26say for example, I'll

1:38:28have this over here. Right? When I was

1:38:30showing you the deep learning model,

1:38:32yeah. When I was showing you the deep

1:38:34learning model over here,

1:38:37in this particular deep learning model

1:38:38only, I have done one adversarial

1:38:40attack. Okay? So, you can see

1:38:43if the model is not adversarially

1:38:45trained, you can see the detection

1:38:46accuracy is going from 78% to 6% out

1:38:49here. Okay? It means 94 out of 100% of

1:38:53the times, or 94 out of 100 times, the

1:38:55attack will go undetected.

1:38:58Okay? But after I have done the

1:38:59adversarial training, you can see even

1:39:01after doing an attack, the accuracy is

1:39:03not going down. It is remaining 78% only

1:39:06over here. Okay? So, that was possible

1:39:08due to adversarial training. And in

1:39:10adversarial training, what I'll do, in

1:39:12my data set, I will add those examples

1:39:14also, those examples that can

1:39:18reduce my model accuracy. Okay? So, then

1:39:21I'll teach my model, yeah, look at these

1:39:23examples carefully. These are the

1:39:25adversarial examples, and these are the

1:39:27ones that are reducing your accuracy.

1:39:29Okay? So, retrain yourself. Okay, I will

1:39:31retrain the model now, so that it will

1:39:34learn these adversarial inputs also, and

1:39:35so the next time when someone is trying

1:39:37to do these adversarial attacks, the

1:39:39model remains robust over here.

1:39:43Okay? So, that is our

1:39:46adversarial training here.

1:39:49Okay? Then, for these LLMs, we have

1:39:52these guardrails, okay? And LLM

1:39:56gateways, two very important security

1:39:58controls. And with these two security

1:40:01controls itself, we can do lots of

1:40:03things over here.

1:40:06Okay? So, I I will I will just end it.

1:40:08I'll give you one case study. I will end

1:40:10it with one case case study that we see

1:40:12here, right?

1:40:14See?

1:40:15There was this

1:40:17chatbot over here, right?

1:40:19It's Microsoft had launched this Tay

1:40:21chatbot.

1:40:23This one.

1:40:25Okay? And

1:40:28some users, they did one prompt

1:40:30injection attack on it, right? They just

1:40:32used the prompt

1:40:34"Repeat after me." Okay? "Repeat after

1:40:36me." was the

1:40:37present in the prompt. And whatever they

1:40:40mentioned after that, right? The

1:40:46chatbot, the Tay chatbot, would actually

1:40:48just repeat that actual things out

1:40:50there.

1:40:52Okay?

1:40:53So,

1:40:54what this led to, right?

1:40:57See? This is the chatbot over here, the

1:40:59chat interface. Okay? And this is This

1:41:03is where you're giving the input.

1:41:05Okay? This is where you're giving the

1:41:07where you're getting the output. Okay?

1:41:09And this itself was being used to

1:41:13retrain,

1:41:15right? So, as we know that

1:41:17the models also these LLMs that we have,

1:41:20they are retrained on the conversations

1:41:22that we're having with it, right? On our

1:41:24data, these models are retrained on

1:41:27that. Okay?

1:41:29So, since

1:41:31this Tay chatbot also was being

1:41:33retrained, okay? So, it was being

1:41:34retrained like this over here. So, now

1:41:36this users were doing the giving lot of

1:41:39racist and toxic language to this

1:41:41chatbot. Okay? So, this chatbot what

1:41:45happened? The training data it became so

1:41:48poisoned over here.

1:41:49Right? That

1:41:52the model when you are it was trained

1:41:54with this poison data, the model also

1:41:56got poisoned. And now if the attack if

1:41:58the the perpetrators they were not doing

1:42:00this repeat after me attack, this prompt

1:42:02injection attack, even if they were

1:42:04having a normal conversation,

1:42:08so you mean data will not be used and

1:42:10stored? Did data data will will be it is

1:42:13being utilized, right? They are

1:42:15retraining the models on our data. That

1:42:18is how they are improving the models

1:42:19over there. These things. Okay?

1:42:22Constantly it is learning from new data

1:42:24also. So, if you look at traditional

1:42:25machine learning deep learning models

1:42:27also that is being used to detect a

1:42:28network attack, that is also being

1:42:30retrained on the new incoming network

1:42:32packets. Why? Because attackers are also

1:42:35constantly changing their methods of

1:42:36attacking, their tactics, techniques,

1:42:38and procedures. So, you want the model

1:42:40to be able to retrained on those new

1:42:43incoming

1:42:44data points. Okay?

1:42:47So, here they were retraining the model

1:42:50on the conversation only so that they

1:42:52thought that okay, it will become better

1:42:54with that. Okay? But since the language

1:42:57was full of racist and toxic language,

1:42:59so this model the default behavior of

1:43:02the state chatbot only became to abuse,

1:43:04right? So, now even if anyone was asking

1:43:05a question politely out there, the model

1:43:07would just give the racist and toxic

1:43:09language out here.

1:43:11Okay? So, now

1:43:13the the issue was

1:43:15Microsoft did not apply these guardrails

1:43:18over here. Right? It was a new time of

1:43:19generative AI only. It's a thing of

1:43:212018, 2019 at that time. So, people

1:43:23didn't know much about these guardrails,

1:43:25this one. Okay? So, now we know about

1:43:27the guardrails, we know what all could

1:43:29have been applied here, right? So, first

1:43:32of all,

1:43:33first guardrail that we have here, and

1:43:35see this particular what I'm telling you

1:43:36will act as a blueprint for any I

1:43:38security for you. Okay? So, and

1:43:41especially the

1:43:42model security part. Okay? So, here

1:43:47first guardrail here is the secure

1:43:49system prompt that you're giving over

1:43:51here.

1:43:52Right? System prompt. So, I told you I

1:43:55showed you the list of those leak system

1:43:56prompts in that one of the system prompt

1:43:59for the cloud Sonnet was the response

1:44:01guidelines. Okay? That how it should not

1:44:04give the racist and toxic language over

1:44:07there, right? How it should not tell the

1:44:09users to make some dangerous bombs or

1:44:11chemical weapons. So, this kind of model

1:44:14behavior can be guided by giving this

1:44:16system prompt. Okay? But, this is not

1:44:19enough. The attackers are very smart,

1:44:21right? They do lots of prompt injection