GenAI Cybersecurity: OWASP Top 10, MITRE ATLAS & API Attacks

In this quick section intro, you’ll get a snapshot of what’s ahead — from understanding what LLMs are and how they work, to exploring how they differ from traditional AI models. You’ll also get a preview of real-world use cases and the key challenges tied to using LLMs responsibly and securely.

In this lecture, we explore what Large Language Models (LLMs) are, how they work, and their role in Natural Language Processing (NLP). Learn about neural networks, key capabilities of LLMs, and popular examples like ChatGPT, Claude, Gemini, and Copilot. This foundational understanding will help you better grasp the power and versatility of LLMs in modern AI applications.

In this lecture, learn how Large Language Models (LLMs) fit into the broader Generative AI landscape. Understand the difference between LLMs and other generative models like DALL·E and Midjourney, and why LLMs play a central role in orchestrating AI-driven content generation.

In this lecture, discover how Large Language Models evolved from basic neural networks to powerful transformer-based systems. Understand key innovations that made modern LLMs possible.

In this lecture, we explore Convolutional Neural Networks (CNNs) and how they recognize patterns in images. Learn how CNNs process labeled photos, detect features, and predict outputs in computer vision tasks.

In this lecture, explore how Recurrent Neural Networks (RNNs) work by processing language sequentially, word by word. Learn why RNNs were important for capturing word order, and understand the limitations that led to the development of Transformer models.

In this lecture, learn how Transformer models solved the limitations of RNNs by enabling parallel processing and scalability. Discover how Large Language Models (LLMs) like GPT were built on transformers and introduced key innovations such as positional encoding, self-attention, and multi-head attention.

In this lecture, learn how transformers use positional encodings to capture the order of words and tokens in a sentence. Understand why positional information is critical for processing language in parallel without losing meaning.

In this lecture, learn how Self-Attention enables transformers to deeply understand the relationships between words within a sentence. Discover how self-attention replaces traditional cross-attention and allows large language models to generate more meaningful and coherent outputs.

Learn how Multi-Head Attention boosts transformer models by allowing them to understand language from multiple perspectives at once. Discover how this innovation enriches context and meaning in large language models.

Review the major innovations in transformer models, including positional encoding, self-attention, and multi-head attention. Understand the foundation behind modern large language models like GPT and Gemini.

Understand how transformer architecture powers foundation models, enabling versatile AI tasks like question answering, information extraction, and more.

Explore real-world categories of LLMs, from copywriting to coding to business ops, and understand the infrastructure companies that support and fine-tune large language models.

In this lecture, we explore the critical challenges and risks that come with Large Language Models (LLMs).

We dive into real-world examples, like the now-famous case where lawyers were fined for submitting hallucinated citations generated by ChatGPT.

You’ll learn why hallucinations happen, why security risks around sensitive data leakage are a growing concern, and how data consent and copyright issues pose legal and ethical challenges.

In this final lecture of Section 1, we summarize everything we’ve learned about Large Language Models (LLMs) so far. From understanding the evolution of neural networks to uncovering the groundbreaking innovations like self-attention and multi-head attention that power today’s LLMs, we bring it all together. We also reflect on the importance of deep technical knowledge for cybersecurity professionals — because you can only secure what you deeply understand. Join me as we wrap up this foundational journey and get ready to move to the next level of exploring LLM architecture and security attack surfaces.

Kick off Section 2 with a quick overview of what’s coming up! In this short lecture, we introduce the internal anatomy of a Large Language Model (LLM) system. You’ll get a roadmap of the key components—like APIs, inference engines, and model layers—and learn how they interact.

In this lecture, we shift focus from the core language model to the full-scale LLM system. You’ll learn how real-world LLMs operate within a larger architecture that includes application layers, integration components, and infrastructure. This zoomed-out perspective is critical for understanding how these systems work—and where their cybersecurity risks lie.

In this lecture, we move beyond just the LLM model and explore the complete system architecture that powers real-world applications. You’ll learn about the critical layers that work together — the Application Layer, Integration Layer, LLM Processing Unit, AI Model, and Infrastructure/Ops Layer. This foundational understanding is crucial for recognizing how modern LLM-based solutions are built, deployed, and managed.

In this lecture, we examine the core of any LLM system — the deployed AI model. Models like GPT, LLaMA, Claude, and Orca represent the intelligence engine responsible for language generation and reasoning. You’ll learn what this layer actually contains, how it works, and why it cannot operate in isolation.

In this lecture, we explore the Infrastructure and Operations layer that powers LLM systems. This layer handles training data, inference workloads, model storage, compute resources, and backend services. We also highlight why this layer is a major cybersecurity concern — containing high-value assets like model weights, logs, and privileged credentials. Understanding this layer is essential for identifying and securing potential attack surfaces in real-world LLM deployments.

In this lecture, explore the LLM Processing Unit — the layer that manages inference-time logic in large language models. Learn how user prompts are processed, routed, and handled with runtime controls like temperature, token limits, and safety filtering. Understand the difference between the AI model and the orchestration logic surrounding it, and why this layer plays a critical role in both performance and cybersecurity. Includes a practical example to explain how temperature affects model creativity.

In this lecture, we explore the Integration Layer of large language model (LLM) systems. Learn how external APIs, SDKs, and plugins connect LLMs to real-world applications. Understand the key difference between internal orchestration in the LLM Processing Unit and external access through the Integration Layer. We also highlight the security risks associated with API exposure, plugin misuse, and unauthorized access — critical knowledge for securing LLM-powered platforms.

In this lecture, explore the Application Layer of large language model (LLM) systems — the user-facing front end that powers web apps, chatbots, plugins, and AI assistants. Understand how this layer collects prompts, delivers AI responses, and shapes the user experience. Learn why the Application Layer is a critical part of the cybersecurity attack surface and how poor design can expose deeper vulnerabilities in LLM-powered platforms.

In this wrap-up lecture, we summarize the layered architecture of a large language model (LLM) system and explain how each layer—from application to infrastructure—plays a unique role in functionality and security. We also discuss why understanding this architecture is essential before diving into attack surfaces and LLM-specific vulnerabilities

In this lecture, you’ll get an overview of what an “attack surface” means in the context of Large Language Models (LLMs). We’ll introduce how the complexity of LLM systems increases the number of exposed components — such as APIs, logs, and prompt inputs — and why this matters from a cybersecurity standpoint. This sets the stage for identifying vulnerabilities and securing LLM deployments effectively.

Learn what an attack surface is, why it’s a foundational concept in cybersecurity, and how APIs, ports, and input fields contribute to potential vulnerabilities. This lecture sets the stage for understanding how LLM systems expand the traditional attack surface.

In this lecture, we explore how modern IT environments naturally expand their attack surfaces through third-party integrations, APIs, mobile apps, and chatbots. Learn how increased connectivity leads to more exposure points — and why this expansion makes securing your systems more complex. This foundational concept sets the stage for understanding risk escalation in public and AI-driven systems.

This lecture explains how public exposure significantly increases cybersecurity risk. We examine how internet-facing APIs, services, and endpoints become prime targets for attackers — and why even a small misconfiguration can lead to major breaches. Understanding this risk is crucial before diving into the complex and highly exposed world of LLMs.

This lecture explains why Large Language Models (LLMs) introduce significantly larger attack surfaces compared to traditional IT systems. Learn how public interfaces, unstructured inputs, dynamic behavior, and tool integrations make LLMs more vulnerable to advanced threats like prompt injection, data leakage, and unauthorized actions. A must-watch before diving into LLM-specific security components.

In this lecture, we introduce a two-perspective analysis of LLM attack surfaces: one from the viewpoint of organizations integrating LLMs into their systems, and the other from the perspective of LLM model providers. Learn how risks emerge differently depending on who owns and operates the system — and why understanding both sides is critical for securing GenAI deployments. This sets the stage for real-world case studies and deeper technical breakdowns.

This lecture explores how organizations that integrate external LLMs, such as through APIs or plugins, significantly expand their attack surfaces. Learn how tools like chatbots and CRMs create new entry points — and why even indirect connections to internal systems can become high-risk. We introduce a real-world case next: the Chevrolet chatbot exploit caused by prompt injection.

This lecture explores a viral prompt injection demonstration by Chris Bakke, who manipulated a Chevrolet dealership chatbot powered by an LLM. By cleverly crafting his input, he made the chatbot respond with legally binding statements about selling a car for $1. While humorous, the case highlights the real risks of user-controlled model behavior — and why guardrails are essential for safe AI deployment.

This lecture explores the architecture and attack surface of LLM model providers, from training data to inference layers. You’ll learn how middleware components like LLM Processing Units, external APIs, and internal databases interact with deployed models — and why this creates a complex, layered security landscape. Sets the stage for case studies on prompt injection, data poisoning, model exposure, and more.

This lecture explores the infamous case of Microsoft Tay — a chatbot that learned from Twitter in real time without moderation. Within 24 hours, it began generating racist and offensive content due to malicious user input. A textbook example of training data poisoning, this case highlights the risks of unsupervised learning and the need for strong guardrails in AI systems.

This lecture explores how a misconfigured ClickHouse database exposed over one million log entries from LLM provider DeepSeek. The exposed data included prompts, API keys, and internal metadata — all accessible without authentication. Learn why inference and log storage must be treated as sensitive infrastructure and secured with the same rigor as user data.

This lecture covers a case study where misconfigured Ollama APIs exposed LLM model servers without authentication. Anyone could query the models freely, leading to unauthorized access, potential data leakage, and model abuse. Learn how to properly secure LLM endpoints like production APIs — with access control, monitoring, and zero-trust principles.

This lecture compares the security responsibilities of LLM model providers and organizations that use their models. From API integration and prompt filtering to infrastructure hardening and log protection, both sides play critical roles. Learn how to distribute risk ownership and avoid security gaps in LLM deployments.

In this final lecture of this Section, we summarize key lessons on LLM attack surfaces from both user organization and provider perspectives. We revisit real-world case studies — including prompt injection, training data poisoning, log exposure, and model API leaks — and extract core security takeaways. We also preview what’s next in Section 4: a quick introduction to the OWASP Top 10 for LLMs, where we map these incidents to industry-standard risk categories. This session bridges architectural understanding with structured threat modeling to prepare you for hands-on security exploration.

This quick lecture introduces Section 4 and the OWASP Top 10 for LLMs — a standardized list of the most critical AI-specific vulnerabilities. Learn why this framework is essential for mapping LLM security risks and how it sets the stage for deeper analysis in upcoming hands-on demos.

This lecture introduces the OWASP Top 10 for LLMs, covering the top AI-specific vulnerabilities like prompt injection, data poisoning, and system prompt leakage. You’ll see how each real-world case study aligns with OWASP risks and how this framework helps standardize LLM threat analysis. Ends with a transition into Section 5’s hands-on API demo.

In this short lecture, we introduce Section 5 of the course — focused on LLM API security. You’ll get an overview of the upcoming live demo featuring an insecure Ollama API setup, along with mitigation steps using NGINX. This session sets the stage for hands-on learning, real-world exposure analysis, and OWASP risk mapping.

This lecture walks through the demo setup used to simulate an insecure LLM API. You’ll learn how an Ollama server is exposed to the internet via a public port and why this mirrors real-world misconfiguration scenarios. We’ll also explain how you can recreate this setup for hands-on practice using Azure or AWS — all with free-tier access.

In this 15-minute hands-on lab, you’ll see how an insecure Ollama API setup can be exploited. We’ll simulate a real-world scenario on a Windows Azure VM—installing Ollama, pulling models, exposing the API port, and demonstrating how attackers can access and manipulate the model without authentication. You’ll also learn about ngrok-based exposures, Shodan results.

This lecture validates the hands-on demo by showing real-world Ollama exposures found via Shodan. You’ll learn how misconfigured LLM APIs are commonly exposed online and explore the behavioral patterns that lead to these mistakes — including reliance on tools like ngrok, poor firewall practices, and the security-skills gap in developer workflows.

This lecture outlines how to protect an exposed Ollama API using NGINX as a reverse proxy. Learn how to enforce authentication, close public ports, and secure inference endpoints in your cloud VM. Includes a detailed walkthrough of the security goals and hands-on mitigation strategy you’ll see applied in the next demo.

In this hands-on lab demo, we’ll mitigate a critical OLAMA API exposure by setting up an NGINX reverse proxy with basic authentication. You’ll learn how to restrict public access to the OLAMA server, configure .htpasswd for access control, and re-route API traffic securely. This is a foundational security step in protecting open-source LLM APIs from unauthorized use, tampering, and OWASP Top 10 LLM threats.

By the end of this demo, you’ll know how to:

• Install and configure NGINX as a secure gateway

• Enforce authentication before model access

• Safely expose OLAMA APIs behind a proxy

• Prepare your setup for future HTTPS upgrades

A lab instruction manual is included in the Resources section to help you replicate this setup in your own environment.

In this lecture, we analyze how a single insecure Ollama API setup can expose your system to multiple OWASP Top 10 vulnerabilities — including prompt injection, data leakage, excessive agency, and supply chain risk. Learn how one overlooked configuration can escalate into a multi-layered threat and why defense-in-depth is essential for securing LLM infrastructure.

In this final lecture of Section 5, we review the key lessons from our hands-on demo of insecure Ollama APIs, explore mitigation using NGINX, and highlight how a single misconfiguration can trigger multiple OWASP Top 10 vulnerabilities. This wrap-up bridges into Section 6, where we begin mapping real-world LLM attacks to the MITRE ATLAS threat intelligence framework.

Get introduced to MITRE ATLAS — the industry-standard framework for mapping adversarial threats against AI systems. In this lecture, you’ll learn how MITRE ATLAS applies to large language models (LLMs), and how it helps security teams classify, analyze, and respond to AI-specific attack patterns with structured threat intelligence.

In this lecture, we explore why the traditional MITRE ATT&CK framework wasn’t enough for AI threats — and why MITRE ATLAS was introduced to map tactics and techniques targeting machine learning systems. Using the high-profile case of OpenAI vs. DeepSeek, we walk through a practical threat intelligence mapping scenario, assuming model distillation as the potential vector. Learn how real-world adversarial AI behavior translates into MITRE ATLAS techniques like ML model access, privilege escalation, and exfiltration.

This short lecture wraps up this Section, highlighting how MITRE ATLAS helps model real-world LLM threats and attacker behaviors. We then preview Section 7 — a hands-on bonus lab experience using PortSwigger’s LLM security challenges, giving learners a chance to apply everything they’ve learned in a guided environment.

In this lecture, we introduce the course and provide a walkthrough of the PortSwigger Web Security Academy lab focused on exploiting LLM APIs. You’ll learn about excessive agency in LLMs and how over-permissioned AI agents can create security risks. A live demo video is included, followed by a guide on how to access and complete this lab yourself using PortSwigger’s free resources.

Watch this hands-on demo where we solve the “Excessive LLM Agency” lab from PortSwigger Web Security Academy. You’ll see how LLM agents can unintentionally invoke sensitive APIs (like password_reset or debug_sql) based on crafted prompts—exposing real-world risks of over-permissioned AI agents. Learn API enumeration, prompt crafting, and exploitation of agent autonomy.

Learn how to sign up for and access PortSwigger’s free browser-based LLM security lab. This slide explains the prerequisites, environment setup, and what skills you’ll build by practicing real-world exploitation techniques in a safe environment.

This short wrap-up lecture recaps the key takeaways from this Section, highlighting how excessive agency in LLM systems can be exploited via backend API access. We also preview next Section, where we’ll discuss future trends, skill paths, and how to stay current in the fast-moving world of LLM cybersecurity.

In this final section, we pivot from vulnerabilities to career growth. This intro sets the stage for a series of focused tips on how to break into LLM cybersecurity, build hands-on skills, and stay current with tools, frameworks, and industry trends.

This lecture reviews the key milestones of the course — including LLM system architecture, attack surfaces, OWASP Top 10 risks, API exposure demos, MITRE ATLAS mapping, and hands-on lab work. It sets the foundation for career tips and next steps in the final part of the course.

In this video, you’ll learn how to identify the security sweet spot where AI and cybersecurity intersect. We’ll explore common attack surfaces like APIs, data leakage paths, and model misuse vectors. Plus, you’ll learn how to think like an attacker — a mindset critical for defending intelligent systems.

In this lesson, you’ll discover how to stay ahead in LLM cybersecurity by aligning with trusted standards like OWASP Top 10 for LLMs and MITRE ATLAS. You’ll also learn how to track research, GitHub projects, and changelogs from core tools like LangChain and Ollama to stay current with the latest vulnerabilities and mitigations.

In this lecture, you’ll learn how to practice LLM cybersecurity skills using hands-on tools and real-world scenarios. From simulating prompt injection in local environments with Ollama to using Burp Suite for API testing, you’ll gain practical experience in attacking and defending AI-powered systems.

Learn how to establish your credibility in LLM cybersecurity by building an online presence. From branding your skills on LinkedIn to blogging on Medium and following key research bodies like OWASP and CRFM, this lecture covers the smartest ways to stand out in the GenAI security space.

In this closing lecture, you’ll get a final motivational nudge to remind you that you don’t need to be a machine learning expert to enter the field of LLM cybersecurity. All it takes is curiosity, a security mindset, and a passion for learning how these systems behave under pressure.

In this final lecture, we reflect on the journey through GenAI and LLM cybersecurity. You’ll hear why foundational understanding is key to securing modern AI systems, and why clarity is your best defense against complexity. Thank you for being part of this deep technical dive.