When (day):

Thu

At:

16:00 - 17:00

Project:

DevSecOps

Watch

Session Video

About this session

In the rapidly evolving landscape of technology, the relationship between developers and application security is paramount. Yet, too often, these two entities operate in silos, leading to overlooked vulnerabilities and missed opportunities for innovation.

In this session, we’ll explore:

The Empathy Connection: Drawing parallels from UX/UI, we’ll delve into understanding the mindset of developers, demonstrating how application security teams can harness humanistic behaviorism to foster enhanced collaboration.
Proactive vs Reactive Security: Utilizing the principles of foresight, we’ll dissect how developers and application security teams can anticipate threats before they manifest. By transitioning from a mindset of ‘fixing what’s broken’ to ‘anticipating and preparing,’ we can usher in a new era of cybersecurity resilience.
Behavioral Economics in Decision Making: A deep dive into how economic decision-making processes can influence secure coding practices. By understanding the inherent biases and motivators for developers, security teams can craft strategies that align with developer objectives, promoting more secure applications.
Enhancing Developer-Application Security Synergy: Drawing from Human-Computer Interaction principles, we’ll investigate the potential touchpoints where developer tools and security protocols intersect, offering streamlined and intuitive processes that bolster security without impeding development.
In our journey to secure the digital frontier, bridging the gap between developers and application security is not just recommended—it’s essential. Through a unique blend of foresight, human-centric design, and behavioral understanding, we can reimagine this relationship, driving innovation and fortifying our digital assets.

Relevant Publication

Media Features

Bridges, Richard (Host). (2023, February). Human Experience with AI . Untold Riches [Podcast]
Bridges, Richard (Host). (2022, August). Think Strongly, Attempt Fearlessly, and Accomplish Masterfully. Untold Riches [Podcast]
Rogofsky, Phil (Host). (2022, May). The Intersection of Psychology & Technology. Fun with Maryland Stem Festival [Podcast]
Marcelle, Carley (Host). (2020, March). Reimagine the Future of Data, Privacy + Security. Matutine Network [Podcast]

Speaking Engagements

IMPACT National Security Conference, Take Your Security Awareness Program from ““Good”” to Great (April 2023)
Bridges Real Estate Summit: Keynote, Brining Foresight into your Real Estate Business (December 2022)
Cybersecurity Summit of Minnesota: Keynote, Hindsight: 20/20 – Bringing Foresight to Cybersecurity (November 2022)
SANS Security Awareness Conference: Keynote, What UI/UX Taught me about improving security awareness (August 2022).
National Cybersecurity Alliance Convene: Scottsdale, What UI/UX Taught me about improving security awareness (July 2022)

Transcript:

Dinis Cruz - 00:00 You.

Jeremy Treadwell - 00:02 Hi.

Dinis Cruz - 00:03 Welcome to this open Security Summit in October 2023. And we have Lewis for the part three or Thread Modeling Qatar, which is a really amazing concept that a lot of people should be following because it’s all about practicing, trying these things and getting a lot of amazing momentum. So, Lewis, over to you.

Jeremy Treadwell - 00:24 Thank you very much, Dennis. So this is Louis Servin. I’m an application security expert, have been working in this area of knowledge for a long time now, over ten years, 15 years almost. Yeah, probably more 15 than ten. I have been doing threat models for such a long time. Oh, I should change the title. Sorry for that. I promise not to. Okay, wait, it’s just a title that I need to change. So again, I have been in threat modeling since two thousand and ten s. I have been threat modeling many different systems as a consultant, as an in house expert across different fields.

Jeremy Treadwell - 01:12 I worked for a long time for an automotive company, and I had the pleasure of being the first person to create a security assessment, a threat model of factory shop floor look at automotive systems, bridge over knowledge for threat modeling from it into the automotive domain. But we’re not discussing that. So that’s just a bit of background. What we’re discussing is threat modeling. And because the topic of this session or this round of the Security Summit was serverless architectures, I decided to, hey, let’s do a serverless architecture threat model. Why not? And as I was preparing this, and I had already chosen the case and I was doing everything, I had a very interesting conversation with Dennis yesterday on a panel were sharing, discussing how to enhance our threat modeling capabilities through something like Chat GPT.

Jeremy Treadwell - 02:14 So I took the bait and I decided to try to do some threat modeling on a serverless architecture and enhance my superpowers or my abilities. No, superpowers enhance my ability to become a superpower by using Chat GPT. So I hope Chat GPT doesn’t make me look like a fool. I did try a very simple rudimentary prompt, but I want to have the real prompt with you. But first of all, let’s discuss what we’re doing. So feel free to interject with a comment or something. I’ll try to keep them open. Let me see, how can I open the comments? Chat here. So if you have any questions, drop a question into the chat and I’ll make sure to answer as fast as I can. All right.

Jeremy Treadwell - 03:15 So threat modeling, I like Brooke Schoenfield’s definition the most, is a technique to identify attacks a system must resist and defenses that will bring it to a desired defensive state. So no chances. Based on business value, we defend more, we defend less. And I chose the title of Kata based on this martial arts practice of performing a set of choreographed movements. If you ever took karate, for example, katas are part of your examination to move from one belt to the other. And the higher up you go in the rank, the more complex the movements are. And catas give you a way to perfect or construct this so called muscle memory. So I’m not inventing the term of CATA for doing technical work. Actually there’s a lot of previous work.

Jeremy Treadwell - 04:17 I like to take the architectural catas from Neil Ford and I think he took the concept from someone doing development catas as well. So I could say at least ten years of catas as a concept for It systems. So I could advise you to look into that if you want to test or flex your engineering muscles. Was I showing the right presentation? So kata. Is this symbol from Japanese? It could be actually two symbols and it means form. Okay, let’s move to the next one threat modeling, just in case no one knows about it or someone doesn’t know about it, is basically a four step process. The first one being understanding what we’re building, modeling the system, the second analyzing the threats, identifying what could go wrong.

Jeremy Treadwell - 05:14 And the most important step is actually the third one, which is mitigation the requirements, the things that need to be added on top to solve the threats that we identified or the vulnerabilities and threats. So one common concept that we see is difficult for many people is to differentiate between a vulnerability and a threat. And I would like to use the opportunity to say a vulnerability is a characteristic of the system. It’s perhaps a negative characteristic, but it’s something that belongs to the system. And a threat vector, an attack is something that exploits this vulnerability for malicious purposes. And cross site scripting is an attack vector, not a vulnerability. Just out of principle, it’s an attack technique. The vulnerability is taking input from an untrusted source and presenting it to the user where it can be interpreted as code.

Jeremy Treadwell - 06:15 And the fourth step, which is validating the outcome, so is the architecture. Is my understanding of the model in step one correct? Are the threats I identified and the vulnerabilities present and reachable? Is the mitigation correct, complete or should I add more or take some? And for this session I took a system from a Tweet. So someone was let me see if I can find the Tweet. Give me 1 second to see if I have I had it open. Let me see if I have the Tweet here. Yeah, I have it here.

Dinis Cruz - 07:01 Can you just quickly paste that content in the chat?

Jeremy Treadwell - 07:05 Of course I can. So I will put it in the slides as fantastic. Yes, it’s there in the chat. Of course. So I took from Raul Hunko this definition of a Qatar. So he’s giving a Qatar for development and he’s saying OK, I need a system to manage books. And he goes on to describe it and it looks very simple like that. So basically I took his definition and made a couple of slides just to have it more present. So what are we building? What is our target of analysis? It’s a bookshop. The bookshop will have books, we’ll have customers, we’ll have orders and payments. Not too difficult. Only authenticated users can make orders. A customer can have many orders, an order can have many order items. An order has one payment, an order item is one book. Easy peasy, right?

Jeremy Treadwell - 08:10 Now, connection to the database he says use an entity framework or Hibernate. Let’s say we use Java, we use Hibernate to connect to the database and map code entities, code classes to tables. So we delegated that and Hibernate will manage the database connections and we will create some rest endpoints. So we have getbooks to get all books postbooks to add a new book which is only accessible to the admin. Delete books and a book ID to delete a book by its ID and it’s the admin get orders, get all orders. It’s something only the admin can do. Post orders is create a new order, get orders with an order ID is to retrieve a specific order.

Jeremy Treadwell - 09:04 So as a customer I can look at my order and see which books were included in the order and I can post a payment process a payment for an order. So that’s what we’ve got. And now we say okay, we have an authentication using JSON web tokens which means there is an identity provider which will allow my customers and my administrators to log in, identify themselves, gain a token to authorize themselves to make calls to the backend system. He says configure serilog. I have no idea what Serilog is, but I will say configure log for J in Java for logging and logs need to be managed by Azure’s AWS log management service and there will be unit tests to test that the endpoints are working and integration tests and that’s about it. I mean, we can look into the database and all that.

Jeremy Treadwell - 10:11 So the point here to start is and the first challenge always is where do I start? So do I jump right into the details or do I want to spend some time looking at the database or at the case in order? And do I have all actors represented in this solution? So what I usually do, let me stop sharing my screen and share my well, I can share it here, so I have to work over there. So let me share my GitHub. If you want to follow along what I will be doing, you can follow on GitHub give me just 1 second to add the link here, let me post it on the chat. So on GitHub LF serving OSS Threat Modeling just lost the chat window. Where’s the chat window? There’s the chat window.

Jeremy Treadwell - 11:49 So if you go to this GitHub you will find so let me put it on the slides to make sure that everyone has it.

Dinis Cruz - 12:02 Yeah, we’re only seeing your slide.

Jeremy Treadwell - 12:03 Yeah, you’re still seeing my slide or not?

Dinis Cruz - 12:07 Yeah, we only seen the slide. Let’s stop here.

Jeremy Treadwell - 12:10 All right, give me a second. I thought you were seeing my so let me add one slide files. So that’s the repository where you can find the exercise we’ll be doing so you can clone it. Let me share my screen then. Not just the file. So share the whole screen. Share. Now I will put my Vs codes on the screen. Can you see my Vs codes?

Dinis Cruz - 13:03 You can?

Jeremy Treadwell - 13:06 Was that a yes or a no?

Dinis Cruz - 13:07 Yeah, you can. All right, perfect.

Jeremy Treadwell - 13:11 Yes. So I started creating a very high level diagram in which we have a customer, an API, and then let’s make the database. So I create there are instructions on the repository to use this JSON file that generates these handy functions I am using. So you can use it. I will create one without images to simplify development. So this is my DB and this is my database. It’s RDS postgresQL. And here we store and serve book and customer data. So basically, what I’m using is the graph language to create graphs, graphical representations of graphs. These elements you see here are called like in any graph nodes. And the links between the nodes are called edges. So basically what will happen is that I have also HTML.

Jeremy Treadwell - 14:38 I could enhance this with the images from AWS, but I don’t want to spend time finding out the image. So I’ll do them without images. Cognito this is my identity provider and this is let mistake, provides pardon? I couldn’t hear you. Well.

Dinis Cruz - 15:22 Just.

Jeremy Treadwell - 15:27 So I’m just trying to model so what I’m doing is I’m trying to model the system as it was described. So I’ll go back. So the first step in thread modeling is understand the system. So model the system. So what we have is the user, right? The customer, not the user. Sorry. So I have.

Dinis Cruz - 15:53 You pushed.

Jeremy Treadwell - 15:58 I’m developing as we speak, right? So the repository I mentioned has the exercises from the previous iterations and has the base file to create these images. And as we speak, I am updating the file. So this is not yet committed. I can commit it right away so that everyone can have it. And then from there we can start working together if you want. So give me 1 second to commit the code. I’ll commit it right away. Customer talks to cognito label equals. So this is provide username password. This goes over Https and that’s it, right? So I will receive provide username and password so we can see that the customer is as I type, the image changes. So I have the first relationship, provide username and password.

Jeremy Treadwell - 17:10 Now I can as a customer talk to the API and I can say label equals make requests. And this is all again protected with Https and JSON web token. So here we have an initial representation of the system in which we have the customer talking to the API, the customer talking to Cognito to get the identity. Now, this doesn’t seem complete, right? There is something missing here. I don’t usually talk to an API, do I? So in order to be able to talk to the API, I need to talk to something else that gives me the access to the API, which sends me to Cognito to authenticate. So let’s say I have a single page application somewhere. Spa. No, not shape. Spa is my single page application. It’s react JavaScript, and this provides the UI for the user.

Jeremy Treadwell - 18:52 So I have a single pledge application, provides a UI for the user. And obviously, the customer needs to talk to the spa label equals get the UI. And this is over Https, but it’s not secured with JSON web token because this is basically me going to Amazon and reaching the page before authentication. Right. The single page application will be the one running on my browser, which will redirect me to Cognito to authenticate if I need to authenticate for buying a book. And then at some point, something has to talk to the database. So the API will be talking to the database, and the API will be talking to the database label equals, let’s say, get write data. It’s unencrypted and admin user password. So that’s how we’re connected with admin user and password.

Jeremy Treadwell - 20:21 I’m obviously creating a vulnerability on purpose so that we can find something easily, right? And this is a very high level view of what could go wrong. Now, there is something missing here. What is missing? So I can make orders. How are books getting into this thing? So I need an admin user, right? So now I have an admin user, or I need an admin user. So I’ll just copy the customer and paint it in a different color to make a difference between or maybe not yet paint it in a different color. But let’s say customer. Now it’s not the customer, but my admin. And now I say admin. Admin is a person internal to the company bookstore admin. The bookstore admin has the function of updating. So let me zoom in so that it’s visible what I’m doing.

Jeremy Treadwell - 21:47 Updating catalogs, or rather than updating updates. That just updates catalogs, checks, orders. And what else was this person supposed to be doing? So this person was also supposed to be doing adding books. So managing books, deleting books, posting and deleting books, getting orders. Those are the three specific functions. So let’s see. Update the book catalog, update book catalog, delete book and checks customer orders. But there was something else missing, right? So what am I missing? So I have get books, post books, delete books, get orders, post orders from the customer, get orders. I’m missing the payment. Post payment. So I need to add a payment provider. Now, this is an architectural decision that’s hard to undo. I can choose to do payment transactions myself, or I can choose to use something like Stripe.

Jeremy Treadwell - 23:19 In my case, I could use a payment provider not to bother with credit card data. So in this case, I will take a system outside of Amazon, say HTML out, so I make it gray to make it visible that it’s out of scope. I will not analyze the gray stuff. I don’t care about the gray stuff. I care about how I integrate with them, of course, but not how they do the business. It’s a different process, vendor requirement, security, whatever, but not in scope for the threat model. So I could say stripe. Stripe. Come on, stripe. It’s a payment provider. I could think of it as a SaaS, as a software, as a service, and it manages payments with credit card returns, a transaction ID. All right, so I have the things that I need to be able to operate now.

Jeremy Treadwell - 25:04 I need to mix them. So I need my API to talk to Stripe and it will send vendor ID, send. What else will I send to it from the API so we can see the line gets generated automatically. I don’t have to generate it myself. API sends the vendor ID, what else does it do? It sends the vendor ID and the amount. Sends a vendor ID and the amount this is an Https connection with an API key. At least that’s how in my mind it works. I have no idea exactly how Stripe works, but I am assuming here that Stripe works this way. Now I have a high level context diagram. I still need to put my admin to talk to my API. And here I will use this label and this label will allow me to manifest the intention of the connection.

Jeremy Treadwell - 26:43 So let’s say the admin updates book catalog and checks orders and this connection is done over HTPs with adjacent web token as the authentication method. So as you can see, the diagram grows as I type the diagram, updates as I need it to update. So I don’t bother myself with creating it in a visual way and moving things around for the reason that it’s a lot easier to have a bit of text, which I can keep in version control, which control the creation of this file, but I can diff them. I can do a lot of things that make life a lot easier. So I can help readability by providing a tab here and making everyone basically use the same length. And it’s a bit easier to see the relationships here, but that’s just a basic readability linting thing.

Jeremy Treadwell - 28:05 And the diagram is relatively simple, so there’s no need to do a lot of work to understand it and we capture the whole scenario. Now, what are things that I could worry about? I could worry about injection, I could worry about many things. Now we can also say, well, this will be hosted in AWS. This will be made up of lambdas. Every API endpoint is a lambda. And I’m not sure if you know Daniel Misler. So let me put on the chat the link to Daniel Misler’s page, 1 second. So there we have it. The other browser 1 second. I’m posting it in the chat. There’s the chat. So here we have the link, and for simplicity, I have summarized what he says on how to use AI, how to shape or what he calls prompt engineering. So he has a seven step process.

Jeremy Treadwell - 29:44 Let me put it over here. Come on. Thank you, so helpful. But I don’t need all your help. One out. So basically he says, okay, tell chat GPT who you are, tell the system what format it will produce, give it the task, give the steps, tell it exactly how you want the output to look for, show it one to five examples of ideal output and tell it what to include or not include. So I was preparing this session by creating this step. So basically what we’re saying is you’re an application security expert, threat modeling a system. You will identify possible vulnerabilities threats in which the vulnerability can be exploited and propose one or several countermeasures to avoid the vulnerability or prevent or detect the malicious action.

Jeremy Treadwell - 30:49 The output will be marked on table with the following fields ID component, Actor, Vulnerability, Threat Scenario and countermeasures. The system is hosted in AWS. The code is managed in GitHub. GitHub actions push the software into production. The system is a single page application on the front end. The system consists of a single page application which the user downloads. The single page application written in react interacts with the rest API calls in JSON users redirected authenticated. The single page application is stored in an S three pocket. Requests are first landing in cloud front, then on the S three pocket we use Cognito as the identity provider and we can do identity federation with Google, Facebook, Office and LinkedIn so that our customers can log in more easily. And on the cloud side, we are fronting our API with AWS’s API gateway.

Jeremy Treadwell - 31:55 Here the JSON web token will be checked and everything will be forwarded to a lambda. The lambda is written in Java or Dell. Lambdas plural are written in Java. They interact with the database using the Hibernate framework RDS postgres instance is the database using what I told you, the database admin to connect to it. We can add to that over an unencrypted connection. The password is configured as an environment variable in the lambda function. The connection is not oh, I already had it. Okay, well, I can’t get it. The connection is not configured to be encrypted, so we can remove that one.

Dinis Cruz - 32:43 This is a great prompt, man. You have to share it. It’s really cool. Especially if we can abstract just a bit right. And just have the bit where I fill this bit here.

Jeremy Treadwell - 32:51 Well, I’m following the recipe from Daniel, right?

Dinis Cruz - 32:55 I know, I know. But it’s a good practice.

Jeremy Treadwell - 32:59 Let’s see what comes out. So I just finished typing this thing. I have no idea what will come out of it. So I hope Chat GPT can run it. And if not, are you using your paid account?

Dinis Cruz - 33:12 Are you using four or 3.5?

Jeremy Treadwell - 33:14 The free one.

Dinis Cruz - 33:16 Okay, so then also put that on the Chat. Then I’ll run it on or push it to GitHub and I’ll run it on four.

Jeremy Treadwell - 33:21 All right, let’s see that. So I’ll commit this to GitHub as well. The application has these user groups. So we have a customer which is external, an admin which is internal logging. We use log for J to produce logs. Logs are captured and logs are captured and stored. Log entries are used for monitoring and debugging purposes. We can say they’re captured and stored in an S three pocket, so disconnected from a lock processing mechanism. So I’m introducing avalibility into the system to see if it’s cut.

Dinis Cruz - 34:12 But can you see how cool it is, what you’re doing right now? You are describing the system, right?

Jeremy Treadwell - 34:17 I am. I mean, it’s the same thing we already had, right?

Dinis Cruz - 34:20 I know, but it’s very different because in the past, remember, even the case were talking yesterday, what would happen is you would spend some time with the engineers, with the developers, with the architects, and you would have this in your head, but you had no place to capture it or persuade. There was no place that you would capture that you get a good reward from it, so you probably wouldn’t capture it. Where now what you can do with this is you can write this, and even if it’s going to the Dev team and architects going, hey, is this correct? There’s already a lot of value in it because we already started to map the things. And then you apply the security analysis on it, but you could also apply performance analysis.

Dinis Cruz - 34:52 You could also say, hey, find me usability problems, find me bottlenecks, find me disaster recovery problems, find me performance. And let’s say fecal points of failure on the architecture from a resilience point of view. So there’s a lot of interesting other questions you can ask, not just the security ones.

Jeremy Treadwell - 35:12 All right, sure, definitely. I agree with you. There’s a lot of value here, right? So I’m just kind of creating my prompt. So let’s see, I have created the persona, I have asked for the format, I have given task, I have given the steps well, I have given the description of the system. I have told it how I want it. Now it tells me it wants five examples or one to five examples. So let’s create one example. Okay, let’s say again the output must be a markdown table and skip the preamble I don’t want it to blah, blah. Skip the preamble. The ID field is a counter used to uniquely identify the fields, vulnerabilities and scenarios. This is an example. So that means I could create something that looks like this. I said I want ID. I want the vulnerability.

Jeremy Treadwell - 36:54 I just mentioned it here. Oh, the component actor, of course, to know where the vulnerability is present. So let’s copy that thing and put it here, the ID, control V, the component actor, the vulnerability, the threat scenario, and the countermeasures. Now I say one, two, 3123-451-2345. This is a markdown table.

Dinis Cruz - 37:34 You don’t need to provide that. He’s going to do it automatically. Don’t worry about it.

Jeremy Treadwell - 37:37 Pardon?

Dinis Cruz - 37:38 He’s going to do the table automatically. You don’t need to do that.

Jeremy Treadwell - 37:41 I know, but I’m following his recipe in which he says, give it an example of ideal output.

Dinis Cruz - 37:46 Right.

Jeremy Treadwell - 37:46 So I’m going to create a very good credible attack vector to have it. So I would say API, insufficient sufficient authorization checks. An attacker abuses the lack of sufficient authorization to elevate his rights to admin and reduce the price of an item. Countermeasures ensure authorization checks are done at the API gateway and individual function. I’m just quoting the ASBs by heart.

Dinis Cruz - 39:02 Yeah. The other thing we could have done eventually this is we could also feed it a large chunk of the mean. We could also say here’s all of the SVS or the SVS that is relevant to this part. Then pick the ones that, in a way, are relevant.

Jeremy Treadwell - 39:22 All right. So I have all of this. Now I will copy paste it into GitHub so that Dennis can run it in parallel. So give me 1 second to put it in GitHub. So, Dennis, here you are. So let me add the file. Explorer OSS Threat Model.

Dinis Cruz - 39:43 You’re going to put that on the main branch or you’re going to put.

Jeremy Treadwell - 39:45 It no, on the bookstore kata. So I have not committed yet. I will commit everything right now.

Dinis Cruz - 39:51 Okay. Yeah. So that branch is not there. Right? So you’re going to put it up.

Jeremy Treadwell - 39:54 It’S not yet there. Yeah, exactly.

Dinis Cruz - 39:55 Cool. All right.

Jeremy Treadwell - 39:56 Good stuff. Prompt markdown. Oh, God. What did it do? What didn’t it copy the thing? All right, try again. Feel better. Control v okay? Not again. Let me see if I can tweak it somehow. Better copy paste. Okay. I have no idea how to copy paste it. Then I can try to put it into okay, let’s see. Control. C I’ll put it into chat GPT. So here’s chat GPT. I have a new session on Chat GPT 3.5. Let’s see if I can put it here. There it is. So now I can select it from here. Post it there. Now it worked. So that means I can now commit and paste. So give me 1 second to commit and paste. Okay, I think I have too many things, so give me one instant to move this to the other side.

Jeremy Treadwell - 41:35 So I will add this into the.

Dinis Cruz - 41:38 By the way, what’s really brilliant about Louis, what you’re doing here is I think you’re really capturing. I would call the essence of threat modeling because I think a lot of people go, oh, threat modeling is know, it’s complicated stuff and all. Like no, like literally, this is almost the state of the art of threat modeling. Right. We’re still figuring out we’re still mapping the things, and then we’re connecting the dots. Right. So this is exactly what happens in the real world.

Jeremy Treadwell - 42:01 Yes. It shouldn’t be hard, right? It has something. Yeah, it has very few things. I could have expected a lot more. It gives me a starting point. Right. So insecure storage and attacker gains unauthorized access to the S Three pockets, storing the single pitch application tampering or data theft. Implement proper s. Three pocket Controls versioning. Not bad. I mean, easy pick. Identity provider, identity spoofing impersonate forging authentication tokens or weaknesses. Enforce MFA. Not bad. Regularly audit and rotate secret keys. Monitor and log authentication events for suspicious activities.

Dinis Cruz - 42:55 Hey.

Jeremy Treadwell - 42:56 Fantastic. So far, so good. An attacker provides a forged or exponent.

Dinis Cruz - 43:00 Have you uploaded the code to your thingy?

Jeremy Treadwell - 43:03 Pardon? No.

Dinis Cruz - 43:07 Just do it quickly because then.

Jeremy Treadwell - 43:08 I was just too eager. But I’ll do it right away. I was just having so much fun. Give me 1 second to do it. So it add bookstore, qatar. All right, I have it. Kit, omit, book push, no trouble. So you could pull now if you want.

Dinis Cruz - 44:13 Okay, well, I’m going to grab it just directly from the website.

Jeremy Treadwell - 44:20 So let’s see. API, insufficient. JSON web token validation.

Dinis Cruz - 44:25 Oh, so you uploaded to Mania.

Jeremy Treadwell - 44:29 I uploaded to Mania. I just had a branch. Not a branch, but a directory for bookstore. I’m developing it myself so I don’t have to do a lot of complicated stuff here in the and it’s all text file, so it’s not too bad. Nothing will break if I corrupt my own repo. Let me export the file. 1 second.

Dinis Cruz - 44:55 No, I got it. It’s already generating.

Jeremy Treadwell - 44:58 All right. I don’t know if you can create a link or make a pull request to see the output. So I’ll just basically I can share my screen or you can share your screen. Even better.

Dinis Cruz - 45:11 Yeah. So I’ll share my screen. You got those. You got nine. All right, let me share my screen. Okay, so where’s mine? Sorry, I was trying to find and what I’ll also do is I’ll show you some of the tricks. Okay? So you should be able to see my screen. Chat GBT.

Jeremy Treadwell - 45:41 Yes.

Dinis Cruz - 45:42 Cool. So you got application system experts. You got this, right? So API endpoints. So they go look. So look. Lambda database vulnerabilities, admin accounts, and encrypted DB connection. Right. So these results better.

Jeremy Treadwell - 45:56 That one was not found by this one.

Dinis Cruz - 45:57 Yeah. So you should be interested to compare both, right? Because. This should be a little bit better, right. In these kind of things, four has a lot more better context and understanding. So, look, use im roles for accessing the database. Single page application, expose the blue token. That’s quite good. There you go. So we give a token look, HTP only. That’s quite good measures, right? Cognitive control over Google. So API, gateway rate limitation, the nail of service, GitHub.

Jeremy Treadwell - 46:35 Oh, you have more things because you have GitHub actions. I didn’t get any.

Dinis Cruz - 46:38 Yeah. Access to git repository. So there’s two ways you can tackle this. So I’ll show you a couple of tricks, right? Lambda, environment variables, API, endpoints logs in S three testing CI pipeline, APIs communication. Right. Where was that vulnerability that we put in? That was by default that you put.

Jeremy Treadwell - 46:59 Unencrypted, unencrypted communication between the lambda and the database. And were using admin database connection between the lambda and the database.

Dinis Cruz - 47:13 Yeah.

Jeremy Treadwell - 47:14 And that wasn’t found. Right. So it does help you get the low hanging fruits.

Dinis Cruz - 47:22 Yeah. But for example, what you can do for here, so we could say, great, can you focus on the connection between what was it with the two elements?

Jeremy Treadwell - 47:32 Between the lambda and the database.

Dinis Cruz - 47:34 Between the lambda and the database. Right. So let’s see. There you go. See? So one of the things that is really worth doing is going bit by bit. So now I picked it up, right? See, lambda to RDS connection, unencrypted connection there in transit could be intercepted by an attacker encryption. You seeing that?

Jeremy Treadwell - 48:06 I see it. I mean, that’s the last step in the recipe from Daniel, right? So that’s why I like Daniel’s recipe so much, because it gives you a very good way to create a prompt and then a very good way to tweak the output. Which is exactly what he’s saying.

Dinis Cruz - 48:22 Right?

Jeremy Treadwell - 48:22 Tweak the output. Focus it on where you want to focus. So it gave us a very good first pass with ten findings on something. But then we can start focusing on the API, we can start focusing on the S three buckets on the log collection or lack of log collection, et cetera.

Dinis Cruz - 48:38 So there’s two ways you can do this, right? You can keep improving the prompt here. One gotcha, which is a hack, is that remember that there’s still a limitation on the prompt size. So eventually, if you keep doing this, you’ll notice that the chattypt will start to lose data because the prompt size is always the same. Right. So one trick that you can do is you can edit this one here. See what I mean? So I’m going to edit this guy and I’m going to say so, for example, you have countermeasures. I could also go, let’s say I’m going to go risk level probability, right? Actually, see, I don’t even need to do that, right? I could also go can you also add to the table above? Risk level probability attack agent and let’s say impact, right, to business.

Dinis Cruz - 49:47 So what’s cool about this, right, is that now it should create a wider table, right? So you could see that expand the table. So increase that. See, look, there you go. See what’s doing?

Jeremy Treadwell - 50:06 I know, looks nice.

Dinis Cruz - 50:12 And then the thing about this, right?

Jeremy Treadwell - 50:14 No, the question here is, okay, you have a probability medium, but a high level risk to understand the likelihood of occurrence, right? Because risk is a vector. So it works two numbers, right? An X and a Y. So we have the probabilities, the Y, but we are missing the X.

Dinis Cruz - 50:36 We do. But I always say that depends on the attack agent, right, or the attack based the attacker. Right? Because whether you have that attacker or not, because if you don’t have internal attackers, then the probability of some of that stuff is happening is quite low, right? Like, for example, this one, not between.

Jeremy Treadwell - 50:57 The Lambda and the database, right?

Dinis Cruz - 50:58 Well, between the Lambda database, if you don’t have internal attackers, it’s going to.

Jeremy Treadwell - 51:04 Have here a high attacker, right?

Dinis Cruz - 51:06 Well, the risk is high. For example, for that one, the Lambda database. By the way, I have a trick. So I created a little let me share my screen again. I have a little JavaScript which I’ll share with you, right? But that basically allows me to do this. Can you see that? So I have a little script that just makes GBT bigger because it’s really annoying that it’s quite limited, right? I have a little JavaScript that I can run on a page that hacks the UI, right, and see how nice this is.

Jeremy Treadwell - 51:49 I know, yeah.

Dinis Cruz - 51:51 So you can see. So in that example, you got the Lambda database, right? Yeah. It’s high, but it’s because you have external attackers inside a threat, right? But then you got cognito. You got that. So I can now say, great, now can you recreate the table? And this is one of the things I always find that when I do thread models, I never like to mix external or internal, like different type of attack agents, right? Basically who you’re defending from, because they will be different. And also your risk metrics should take that into account. So I prefer sometimes to create the same thread model, but from different points of view because everything becomes a lot more realistic.

Dinis Cruz - 52:40 Because if the business is going okay, so I’m really worried about attacks from the outside, then you go like this, because sometimes you might go, well, the insider thread can do a whole bunch of other stuff. So this is not, yes, you could do this, but it probably can do even more things, or it depends where you are. So I could say, look, now can you recreate the table only from the point of view? Right? The table only from the point of view of external attackers with no to, I say low to mid level skills. Right. So again, I’m not saying state actors and stuff like that. Now, it should be able to refocus a lot of this, what’s it called, the high, mediums and lows should now be reclassified on those. Because now that’s what I would expect.

Jeremy Treadwell - 53:49 It included new things, didn’t it?

Dinis Cruz - 53:51 Yeah, well, it was out there in secured direct object references. Was that there?

Jeremy Treadwell - 53:55 No, I see either number three or four. Number two or three, yeah.

Dinis Cruz - 54:01 There you go. From an external attacker. And we can classify this. So we can also say what is our classification for this? What’s our classification from this? So that’s quite nice. Right. Because you remove actually, a bunch of the ones that are not immediately exposed from those skills. So it’s a lot more focused. S, Cloud, Front, Cognito API, Gateway API expose sensitive data and GitHub if it gets exposed. So this is pretty cool. Another thing you can do here, you could say, actually we can go back to the first one and again we can edit. So I tend to do this quite a lot sometimes edit the original prompt and you could say here, can you map in fact, this should actually work. Can you see what I’m asking for?

Jeremy Treadwell - 55:04 Stride category.

Dinis Cruz - 55:06 Yeah. So this is where the T from the GPT part is super critical because it should now go to Stride and it should give it look.

Jeremy Treadwell - 55:19 See.

Dinis Cruz - 55:22 How cool is that? It understands what Stride means, right?

Jeremy Treadwell - 55:30 I know, but do you know how many times I have spent half an hours of my life discussing whether something falls within this or that category of Stride because it’s such a bad yeah, but that’s a threat category that I tend to skip it. But I see the value in this. I mean, I would rather map it to CAPEC, which is like a lot better.

Dinis Cruz - 55:54 Okay. How do you spell that?

Jeremy Treadwell - 55:56 C-A-P-E-C. CAPEC.

Dinis Cruz - 55:58 C-A-P-E. Like that? Yeah. Okay. And by the way, the script that I was talking about sorry, just because I got some BOOKMARKS, I don’t want to share my screen.

Jeremy Treadwell - 56:16 I’m recording while you get there, we have a cool mel drews is asking any predictions or thoughts on how enterprises could make use of similar techniques now or in the future? Considering with consideration to concerns for sending organizational applications, security data to some AI, I think that’s part of the risk equation. Right. So having models in hand.

Dinis Cruz - 56:44 But that’s already addressed already. Because if you look at Open API, Azure Open API, and if you look, for example, at Bedrock, put this way, if your organization is not using AWS, then you have bigger problems. Right? For example, if you’re using AWS, then AWS already has a mode where you can use Cloud Two that is locked. And if you’re using Google Cloud, it already has Vertex that is locked. And if you’re using Azure also, right?

Jeremy Treadwell - 57:12 And if you have some sort of chat internal thing.

Dinis Cruz - 57:17 I think there’s a couple of variations, but it already completely addresses, I think, Mel, that point. Right. And there’s an interesting argument here where the better we create the prompts, the better we create the materials, the less powerful model you want. In fact, if you really want to go private, use Llama Two, right? And llama two is already pretty good. And again, and you can run that completely in your environment, your data center, your local machine, and there’s different models of Llama. I haven’t fully run it, but some guys I was talking to, it’s about between three and ten grand a month. So depends on what size of the enterprise. It’s already a reasonable cost and it’s going to come down. Right. Again, depends on usage. Right.

Dinis Cruz - 58:01 If you have one person using, maybe it’s a lot, but if you have a lot of developers and if you really have those requirements not to send things very widely, then that makes a big difference. So, yeah, let me just share my screen to see the output here.

Jeremy Treadwell - 58:15 Cool.

Dinis Cruz - 58:16 Can you see? Yeah. So there you go. Right? Is that better for you?

Jeremy Treadwell - 58:23 So KPEC is really the better mapping thing rather than strife.

Dinis Cruz - 58:28 Yeah. Now, the thing about this is that this is what the guy that was doing the attack gen thing is that some of these sometimes JBT does get these tend to get more the IDs wrong than these mappings here. So what you can do is you can feed it to it, right? So you can actually feed for example.

Jeremy Treadwell - 58:44 You can ask it to put the link, right? Now, if it’s linked, you can just click and check the.

Dinis Cruz - 58:51 That’S the one, right? Common attack pants classification, right?

Jeremy Treadwell - 58:54 Yeah.

Dinis Cruz - 58:57 But the other thing you can do with this is you could also say, see, look, I don’t know if you’ve seen this, so I could also take this and I can basically go a new one, right? And basically I can use the advanced data analytics, right? And I can then say I was actually playing with this and you can basically say, can you create a visualization right of this application from a thread model point of view? So in principle, this is the one that has the code, right? So basically you could see that, first of all, it’s going to create the table with probably a bunch of issues. In fact, I could have probably given the other issues. Right, see, so it’s doing that, but this is the one that has the advanced data analytics. Have you used this before?

Jeremy Treadwell - 59:57 No, I don’t have a paid subscription. Yeah, my company is not yet on Chat GPT. I’m actually supposed to be doing the analysis for the risk assessment, but when I’m done with that, we might be able to use it. But I mean, right now I’m on my private account just trying things that have no relationship to my company, right?

Dinis Cruz - 01:00:17 It’s really worth doing this because this part here, it is quite impressive and sorry, I’m just saying.

Jeremy Treadwell - 01:00:44 I assume based on the description, you could also ask it to stop right where we told it to create a table and tell him or tell it to create mermaid diagram.

Dinis Cruz - 01:01:00 Exactly. That’s what I’m going to do now. So you see, look, see the risk levels? Okay, so basically it’s done this bit, right? You got this thing and now you can go, great. Can you create, let’s say, graph visualization of this data, namely the data application should be able to now this is the one that has so you can see that now he’s mapping, right? He interacts single page authentication handed to government success authentication. The user receives a JWT token. API requests are out to the API gateway. And my point with all this is that we should still have, you know, this should still be verified, right? But this is the one that actually erites the code. See? So he’s using Matplotlib with network X. He’s going to create a graph. See, look, there’s the nodes, there’s the edges.

Dinis Cruz - 01:02:15 Can you see user, he’s adding the edges. He’s drawing the network, getting some labels to it. Access request returns. And then he’s drawing the network. And then when he raises bugs like this and then sometimes it’s funny, right? He goes, oh, it doesn’t exist, I should be using this one. Why don’t you do that in the first place? But the other thing that is sometimes really cool is if you have APIs that can do this, you could also ask it for the code to execute in your end or the functions to execute in your end. And you could also give it code samples. But the logic is that it just gives you that acceleration, right? It just gives you that extra mappings and there you go, right? So there’s one visualization.

Jeremy Treadwell - 01:03:03 I mean, it’s something.

Dinis Cruz - 01:03:08 So you can actually map it up, right? Now, the one I was playing around before is like, for example, you could also ask to do a plan to ML. So you can say, hey, give me the plan to ML code for this. And in your example, remember that you already had some of those modules. So you can say, hey, here’s the building blocks that I want. Now use those building blocks to do it. I think once we package this a little bit more, it’s going to be super powerful. And remember that it’s also about helping, let’s say you or the more expert threat model person to review the materials. Imagine you being on the receiving end of this, but also you now communicate in prompts.

Dinis Cruz - 01:03:48 So you can capture your knowledge in prompts, or are you making the original prompts better, which is that much more efficient?

Jeremy Treadwell - 01:03:59 Yeah, certainly this has just so much potential, right? It’s incredible what can be achieved.

Dinis Cruz - 01:04:09 Actually, the other one that you can do. Let me show you. This is while you were doing that. Let me see if I can load it up here. I basically asked? Yeah, it’s kind of okay, give me a second. See if this works. You kind of pause in the middle. But see, the other thing you can do, check this out, is you can take your diagram. Look at this. So what I did is I took this one, see where you were literally a screenshot of your stuff and say, can you write a thread model for this application? And you wrote this, right.

Jeremy Treadwell - 01:05:00 Not bad.

Dinis Cruz - 01:05:01 Phishing attacks, denel of service attacks, API vulnerabilities, this. So it’s not bad, right? Maps the thing that you have there. Yeah, actually, the other trick that you can do, which I also think is insanely powerful, is I was trying to recreate the whole thing. The other thing you can do that is also crazy powerful is we could also take that table. See, look, I think this is really cool. So I can take this table and look, just copy this, right? So I’m just going to copy literally this stuff and I’m going to go, okay, I’m literally just going to paste that whole thing and say, hi, can you review find blind spots in the thread model below. And you can literally just paste it like that.

Dinis Cruz - 01:06:06 See how I’m not even formatting it correctly, but in a way Chat DBT can interpret that way better than we. So this is what I think is really interesting is when you do something and you go, hey, can you find blind spots? Can you look at other things that we could be looking at?

Jeremy Treadwell - 01:06:28 Yeah, certainly. This has a lot of potential, right. It’s basically the beginning of how to do threat models enhanced by AI. Looking at this from the business perspective, I am creating a security champions community right now. And I don’t have a lot of security expertise in the company to the threat models. Brilliantly. They will get there sometime. But this can just turbocharge everyone’s abilities exactly.

Dinis Cruz - 01:07:05 Right. And also, I think you could also drive a little bit the whole thing of thinking of and teaching and getting people to use OpenAI right. And Chat GPT because you’re also teaching them. Because look at this. So I can now can you create a business analysis to provide to a business and project owner to make the business case to spend time addressing these issues? So now, look, create a compelling business analysis, right? And there you go. Right, okay. What is actually giving me okay, so this is a structure approach. This is cool. Look. So you could see introduction, executive summary, introduction, analysis, financial analysis. This is not bad. So it gives you the thing now. Great. So now can you write one adding can you write one based on the data provider? Right? So now it will draft a business one. Right?

Dinis Cruz - 01:08:23 See look. So now it’s going to give you a text that you can rewrite and you can basically apply. So that’s blah and the grand through landscape or threat model in five vulnerabilities, right? Look, see, bam, bang. And if you even know better who you’re going for, you can start to.

Jeremy Treadwell - 01:08:44 Really you can describe the person’s role.

Dinis Cruz - 01:08:46 Describe the person, right? So you can say, Great, but that is too long. Can you give me the same with max one paragraph? Right? Yeah. But now.

Jeremy Treadwell - 01:09:14 This is the kind of feedback that you could give a human and be punched for, right?

Dinis Cruz - 01:09:19 Yeah.

Jeremy Treadwell - 01:09:20 Someone could have spent hours doing this and then you say, come back with one.

Dinis Cruz - 01:09:24 Exactly, right, but that’s the thing. In the past, this is my point. My point is, in the past, this would be really hard to do, right, because we do all this stuff and it’s super painful to do, right, and now you can go, okay, can you write that one paragraph for somebody that has no idea what cybersecurity is? Right? So you can now start to trick it, right? And then you could see that the language changed, right?

Jeremy Treadwell - 01:10:03 I agree, world of Chat GPT is fantastic, but we definitely need to check the output, right? So it is going to write for you very easily or very long text. Now you need to check them and make sure that they actually match what you want to say.

Dinis Cruz - 01:10:22 But that’s the point. The point is that I think where people get in touch with you wrong is that they shortcutting the last step. And the point is that it’s not about shortcutting the last step. The last step is still there, or even the intermediate steps. The point is that the recreation of the materials is now super easy. And like you said, you can go from big to small to change, to multilanguages, to adding mix or say, oh, he’s the language that we’re supposed to do. Or, by the way, this analysis, or he’s a project report or whatever, you can add lots of bits of data in ways that in the past would make it very expensive. Because you have to go back and rewrite and think about these things. Right?

Dinis Cruz - 01:11:01 So even little tweaks, you can say, oh, don’t use language like this, don’t use words like this, use words like that. And then you can rewrite it. And that’s the keys. The feedback loop now of creative materials is actually a lot more, which is really powerful. And if you look at the CATA stuff, I think it’s really worth exploring that onboarding, that practicing. I like the fact that we can now create even games on this. In fact, look at this, right? So actually, let me give a really cool example. If I start from here, right? So let me take it from that guy there.

Jeremy Treadwell - 01:11:43 You’re not sharing your screen, just showing something.

Dinis Cruz - 01:11:46 Sorry. So if I start from. Look at this example. I’m going to take that example of where were actually capex, see where were, where we got that list, right? Okay. So can you now create a game and the set of questions and answers to teach a new developer, right. How to do thread modeling? Threat with a t. Yeah, sorry. Yeah. Although you actually would pick it up.

Jeremy Treadwell - 01:12:22 Thread modeling, I’m not sure. Threat and thread.

Dinis Cruz - 01:12:25 Okay. But it understands the context. That’s the other cool thing. There you go. But it’s fine, it’s better if you spell it properly, right? But now, Silo, can you create a game and set a question and answers to teach a new developer how to do threat modeling for the app above, right? So give it a cool name and use emojis. I usually go with taste, or else you can go quite off piece, right? So call Cyberslu thread defenders to engage. Right? There you go. Look at this. Okay, so what’s cool about this is you. Okay, so it’s actually giving the answer, right?

Jeremy Treadwell - 01:13:09 Can you see?

Dinis Cruz - 01:13:10 But that’s cool, right? Because it already gives you this, right? So you have the thing. But can you see how really cool is to do a game based on that particular application that is relevant to that particular thing? See what kind of risks associated with storing data like DB nine passwords in environments like lambda. Now, this is very relevant to whoever, to the developer that wrote this.

Jeremy Treadwell - 01:13:41 Sorry, I see. What would be fantastic is I’ll put the output of chat TPT 3.5 or. On the repo. If you could also add the output that you have been generating.

Dinis Cruz - 01:13:58 Yeah.

Jeremy Treadwell - 01:14:00 We could then create a branch on the Open Security Summit to keep these things. Also, this is a place to start, right?

Dinis Cruz - 01:14:08 Yeah. Look at the difference. See what happens now. So look at the power, right? See, look, see what happened. I could have gave him this, right? So I could have say, hey, based on this, now run the game. So that will be the prompt, right? So you could see that. Now you got this. Look. Welcome to Cybersfluid. Your first misidentified, right? Walk, comment, blah, blah. And then you can go like this. And then you type the answer and then he’s going to give me a response, right, based on correct answer, blah, blah, fantastic. Identified. But the other thing you can do, which is quite funny, right, is I can come in here, right, and I can say, can you run for me? Can you provide highly sarcastic answers, for example, right? So it’s not going to be for everybody, right?

Dinis Cruz - 01:15:05 And comments to the user, to the player responses, right? So I think this is not for everybody. But look, so now I’m going to say let’s step to the world of this, right, with a good attitude, right? So you can see it’s going to change a little bit, right? Oh, a new cybersecurity on this. Let’s see if you can spot a vulnerability better than it can spot the laser dot, right? So let’s say I’m doing the wrong one and now he’s going to go it should provide a nice sort of really? That’s a cute guess effect. That was chocolate teapot. But don’t worry, there’s more fun ahead. Do you know what I mean? But I think that this makes a difference, right? And you can even put into context. I’ll run this from the point of view of a retailer or a finance institution.

Jeremy Treadwell - 01:15:58 You know what, I’ll try this next week. I have my Cybersecurity Champions meeting next week. I’ll provide an illustration of the system and I’ll let them play the game. Let’s see what happens.

Dinis Cruz - 01:16:09 So what I’m doing with the Cyber Boardroom, right, is to try to create an environment that wraps this so that you don’t see the middle. For example, I’m using Chat GBT behind the scenes, but I, for example, just give you the options and it will give you ABC, and then you click know, so it’s almost like wrap a little UI from a board member point of view, but make it much more interactive, much more kind of nicer. But the engine behind is sort of this logic of Chat GPT, which is really cool, right?

Jeremy Treadwell - 01:16:39 It is definitely good.

Dinis Cruz - 01:16:41 All right, man, this has been great.

Jeremy Treadwell - 01:16:45 It’s definitely a very different threat modeling session than I intended yesterday. But I think after discussing on scaling threat modeling, I really wanted to try a very good prompt for just flexing the prompt muscles and you could see how it works.

Dinis Cruz - 01:17:08 Can you see how a lot more effective you are when you’re finding blind spots versus creating from scratch? And I always felt that the problem with most people doing threat modeling was the start, right? It was that, how do I start? Where do I go? What’s the first? And then you go to stride, and it’s like guessing where even if jump. Hey, thanks, guys. Nice comments there. But look, just what we’re doing here, literally, me and Lewis, is what we do in threat model. This is what we’ve been doing for the last ten years, right? Literally, we’ve been ad living figuring this out. Even the stuff that Louis now is already very comfortable doing. The plan to ML, I remember when we start using that, I was like, well, this is really cool. We can now visualize the threat model.

Dinis Cruz - 01:17:57 So I think what’s nice about our industry is still there’s still so much ground to evolve, and practices like threat model are still very archaic if you compared to what they should be. So there’s great space for innovation, which is really cool.

Jeremy Treadwell - 01:18:13 It is. It’s been a fun session. I know. Next open security summit. Let’s do a different Qatar.

Dinis Cruz - 01:18:23 Oh, yeah. By then we’re going to have a bot. We’re going to have the Open Security Summit bot with us, right. Providing help and enhancing things like this. Right.

Jeremy Treadwell - 01:18:33 I mean, it took me about the same time it would have taken me to draw the thing and walk through the case with creating the prompt. Right. So the prompt was a lot of text and a lot of tweaking and a lot of following the rules, but it’s given us a good result. So we could perhaps next time start with a bath prompt, improve the prompt as we go to see how it improves.

Dinis Cruz - 01:18:59 Well, and to share it. Right.

Jeremy Treadwell - 01:19:02 The idea is obviously to share it in such a session and have the iterations the prompts on a markdown file or something to keep it.

Dinis Cruz - 01:19:10 Yeah. And that’s one of the other bits I’m really looking at also, and playing around with the boardroom bot is I’m creating three bots, right? So I’ll go in detail the session, but I’m creating three bots, and each bot basically contains Minerva. One is Athena, which does the board, the advice. One is Odin, which actually does the backend stuff, and one is Minerva, who does the business stuff. So you kind of have a mix of multiple bots playing together, which know when it works, it’s going to be pretty cool. All right, man, this is good stuff. I think it’s been a pretty good session. Man, this has been a really nice session.

Jeremy Treadwell - 01:19:46 It was really fun. Thank you for your interaction. I mean, it was great to see how you use it. I saw your reading list for the summer and I was like, envious of where you find the time to read so much.

Dinis Cruz - 01:20:01 Yeah, but it’s like I was telling point the point of having those books, although I drove to Portugal so I could bring it, so my wife didn’t complain that I brought a bag just for books. Right. But I felt that I had the books I needed to read in the sequence that I wanted to read. And what that allowed me to do is allow me to go quite deep. And then one of the things I realized was that the action was not in the models. The more depth I went, the more I realized the history of it and where were and the innovations, and the more I understood how GPT worked, the more I realized that the action is not in the models is in the prompts, is in, maybe even be the training of it. Right. But that’s where I became quite interesting.

Dinis Cruz - 01:20:39 But I was able to almost read the book in sequence that I needed to. So I needed the 30 books so I didn’t read.

Jeremy Treadwell - 01:20:46 I mean, I remember your list was like ten books. So out of the ten, which one was the best?

Dinis Cruz - 01:20:52 There was a couple. I think the O’Reilly GPT transformers was really good.

Jeremy Treadwell - 01:20:56 All right, so I’ll get that one.

Dinis Cruz - 01:21:00 I think it’s that one. Yeah.

Jeremy Treadwell - 01:21:02 I’m using you as my chat GPT.

Dinis Cruz - 01:21:04 Yeah, you’re.

Jeremy Treadwell - 01:21:06 My dean GPT. And I cut the ten books to one. That’s good enough for me.