Practical AI Security: Attacks, Defenses, and Applications (on-demand)

Practical AI Security: Attacks, Defenses, and Applications takes you from the foundations of machine learning to advanced security practices involving Generative AI and Large Language Models. Through hands-on labs, you'll train models, build LLM-powered applications, and execute real-world red team attack scenarios. You'll master both the offensive and defensive sides of AI security, from creating Automated Pentesting Agents and exploiting Prompt Injection vulnerabilities to implementing MCP Server attacks, conducting AI threat modeling, performing secure code reviews, deploying AI Gateways for protection, and designing robust guardrails. Along the way, you'll work with industry-standard tools including Hugging Face, LangChain, OpenAI APIs, Anthropic's Claude and MCP, and learn enterprise frameworks such as Google's SAIF (Secure AI Framework). By the end, you'll walk away with practical exploit code, reusable defensive tools, comprehensive security checklists, and the expertise to both attack and defend AI systems in production environments, making you a complete AI security professional capable of operating across the entire security lifecycle.

Target audience

This course is ideal for anyone interested in learning about the application of AI in cybersecurity. Level Beginner / Intermediate. 

Prerequisites

To successfully participate in this course, attendees should possess the following:

  • Working knowledge of cybersecurity and pentesting fundamentals
  • Understanding of fundamental programming concepts and looping structures in at least one
  • higher-level language used in machine learning (eg: Python, or similar)

 

 Se alle kurs innen dette fagområdet

 

Course content:

Key Objectives

  • Understand the core concepts distinguishing AI, Machine Learning, and LLMs, including supervised vs unsupervised learning, neural networks, Generative AI, diffusion models, and the complete ML model training lifecycle from data preprocessing to deployment.
  • Master the fundamentals of Large Language Models, including Transformer architecture, tokenization mechanisms (BPE), context windows, embeddings, and the differences between foundational and fine-tuned models like GPT vs BERT architectures.
  • Become proficient in Prompt Engineering techniques including system vs user prompts, prompt templates, leaked system prompts analysis, and controlling model output via sampling parameters (Temperature, Top-k, Top-p) for security-focused workflows like threat modeling assistants.
  • Learn to use essential AI development tools including Hugging Face Transformers, LangChain (with memory and tool integration), LlamaIndex (multi-file processing), OpenWebUI for local LLM deployment, vector databases like FAISS for RAG implementations, and fine-tuning workflows.
  • Build and deploy production-ready AI applications, including custom RAG (Retrieval-Augmented Generation) systems with vector storage, conversational agents with short and long-term memory, AI-powered security tools with proper rule-based and advanced guardrails, and FastAPI-based scanners.
  • Master Model Context Protocol (MCP) servers for integrating AI with security tools to understand MCP vs traditional connectors, build custom MCP servers, and leverage them for reverse engineering, Mobile malware analysis, and automated penetration testing workflows. Configure MCP with Cursor and Claude for enhanced AI-assisted security research.
  • Develop Offensive AI capabilities, including building autonomous AI agents and workflows for vulnerability scanning, CVE finding, reconnaissance, IAM policy analysis, threat intelligence gathering, and exploit development assistance using frameworks like LangChain.
  • Execute advanced attacks against AI systems, including Prompt Injection variants (direct, indirect, multimodal attacks on CV screeners, meeting summarizers, image analyzers), jailbreaking techniques, data exfiltration through prompt manipulation, and exploiting MCP server vulnerabilities (Confused Deputy attacks, information disclosure, bruteforcing, arbitrary file read/write).
  • Implement Defensive AI strategies, including securing AI-powered applications against prompt injection, analyzing vulnerabilities in "vibe-coded" AI-generated applications, securing MCP servers with proper authentication and authorization, and applying pre-launch security checklists for AI-assisted apps.
  • Deploy and configure AI Gateways to secure production LLM applications and learn to migrate existing apps behind AI Gateways, implement multi-layered guardrails for input/output validation, configure rate limiting policies, and leverage analytics and comprehensive logging for monitoring, compliance, and cost optimization.
  • Master AI-powered Threat Modeling using STRIDE methodology and understand the engineering logic of systematic threat modeling, leverage LLMs to identify threats across Spoofing, Tampering, Repudiation, Information Disclosure, Denial of Service, and Elevation of Privilege categories, and develop practical mitigations with AI assistance.
  • Threat Modeling AI Applications based on STRIDE, MAESTRO's seven-layer architecture, MITRE ATLAS, and OWASP MAS applied to a real-world AI Meeting Scribe.
  • AI Governance & Compliance accross EU, US and India.
  • Fine-Tuning LLMs for Security with QLoRA and Unsloth hands-on
  • Apply AI to enhance Security Operations and Reverse Engineering workflows using Fabric AI for knowledge mining, log parsing, email header analysis, threat intelligence processing, video knowledge extraction, breaking language barriers in security research, and integrating AI into tools like Ghidra and JADX for automated malware analysis.
  • Understand and implement enterprise AI security frameworks, including comprehensive coverage of Google's Secure AI Framework (SAIF) with all 14 security risks (data poisoning, unauthorized training data, model tampering, prompt injection, model evasion, sensitive data disclosure, etc.).
  • Debug, intercept, and secure MCP implementations using MCP Inspector for debugging, Burp Suite for traffic interception and modification, and apply the comprehensive MCP Server Security Cheatsheet for identifying and remediating common vulnerabilities in both custom and third-party MCP servers.
  • Secure AI supply chains by pinning dependencies, verifying model signatures, understanding format risks, and detecting tampering or backdoors.
  • Earn the Certified AI Security Researcher (CAISR) certification by demonstrating mastery across all course modules from foundational AI/ML concepts through advanced offensive and defensive AI security techniques in real-world scenarios.

 

01 Course Preview

  • Practical AI Security: Attacks, Defenses, and Applications - Trailer (10:21)

02 Welcome to the course

  • Getting Started
  • Ethical Use Note
  • About The Labs
  • Setting up the Environment
  • Ask the Instructors
  • Glossary
  • Rise of the Transformers: Attention is All You Need (02:37)
  • Homework

03 Foundations of AI Security (ML, LLMs etc)

  • About this Module
  • Glossary
  • Module Labs
  • Intro to Artificial Intelligence (06:41)
  • AI vs ML vs LLM (05:08)
  • Quiz 3.1
  • Quiz 3.2: AI vs ML vs LLM
  • Supervised vs Unsupervised Learning (02:16)
  • Lab: Creating a Spam Detector (02:41)
  • Lab: Customer Clustering (02:21)
  • Reinforcement Learning (09:28)
  • Quiz 3.3
  • Neural Networks Explained (08:06)
  • Generative AI (01:44)
  • Lab: Generative AI Example (03:00)
  • Diffusion Models (01:26)
  • Lab: Using StableDiffusion for Image Generation (01:50)
  • Quiz 3.4
  • Homework

04 Inside Large Language Models (LLMs)

  • About this Module
  • Glossary
  • Module Labs
  • What is a Large Language Model (LLM)? (05:59)
  • Quiz 4.1
  • Lab: ChatGPT and Grok. GPT or BERT?
  • Solution: ChatGPT and Grok. GPT or BERT?
  • From Words to Tokens: How Text is Split (03:18)
  • Lab: Looking at a Simple Tokenizer (03:09)
  • Foundational and Fine-Tuned Models (02:01)
  • Embeddings: Giving Meaning to Numbers (03:35)
  • Lab: Visualizing Vectors and Embeddings (03:00)
  • The Context Window: How Much Models Can Remember? (02:17)
  • Quiz 4.2
  • Understanding Temperature and Parameters (05:07)
  • top_p and top_k parameters
  • Quiz 4.3
  • Homework

05 Building with Large Language Models (LLMs)

  • About this Module
  • Glossary
  • Module Labs
  • The Transformers Library (04:47)
  • News Summarizer Using Pipelines (03:25)
  • Quiz 5.1
  • Lab: Building a Q&A Tool (00:42)
  • Lab: Building Unsafe Text Detector (02:01)
  • Running LLMs Locally (06:41)
  • Running LLMs Using OpenWebUI (04:19)
  • Quiz 5.2
  • GuardRails Explained (10:05)
  • Lab: Simple Rule Based GuardRail (04:00)
  • Homework
  • Quick Feedback Form

06 Building AI Pipelines with RAG, LangChain & LlamaIndex

  • About this Module
  • Glossary
  • Module Labs
  • Retrieval Augmented Generation (RAG) and Vector Storage (10:45)
  • Lab: RAG Demo Using SentenceTransformers and FAISS (02:45)
  • Quiz 6.1
  • Generating and Using API Keys (03:06)
  • Lab: Generating and Validating API Keys for LLMs (04:02)
  • Quiz 6.2
  • Intro to LangChain (05:33)
  • Lab: LangChain Demo with Short Term Memory (05:10)
  • Quiz 6.3
  • Lab: LangChain Demo with Memory and Custom Prompt (02:03)
  • Lab: LangChain Demo with Memory, Custom Prompt and External Tool Connection (08:59)
  • Quiz 6.4
  • LlamaIndex Explained (08:20)
  • Lab: LlamaIndex Demo with Multiple File Types (07:33)
  • Quiz 6.5
  • Homework

07 Fine Tuning Models

  • About this Module
  • Glossary
  • Fine Tuning LLMs Explained (08:17)
  • Fine Tuning Process (07:17)
  • Lab: Instruction Tuning LLM for Phishing Classification
  • Lab: Fine-Tuning for MITRE ATT&CK Log Classification
  • Quiz 7.1
  • Homework

08 Practical Prompt Engineering for Security Workflows

  • About this Module
  • Glossary
  • Prompt Engineering Explained (06:08)
  • System vs User Prompts (06:41)
  • Quiz 8.1
  • Leaked System Prompts
  • Reusing Prompts
  • Prompt Templates (02:37)
  • Quiz 8.2
  • Prompt Engineering Resources
  • Quiz 8.3
  • Lab: Creating a Threat Model Assistant (05:43)
  • Lab: Adding UI to Threat Model Assistant (02:37)
  • Quiz 8.4
  • Homework

09 MCP Servers for AI and Security Tool Integration

  • About this Module
  • Glossary
  • Intro to MCP Servers (10:22)
  • Connectors vs MCP Servers
  • Quiz 9.1
  • Lab: Reverse Engineering Using MCP Servers (10:25)
  • Lab: Android Malware Analysis using MCP Servers (08:40)
  • Lab: Automated Pentesting using MCP Servers (08:14)
  • Quiz 9.2
  • AI Agents vs MCP Servers
  • Creating Your First MCP Server (07:51)
  • Quiz 9.3
  • Homework

10 AI Agents and Workflows for Security Automation

  • About this Module
  • Glossary
  • Understanding AI Agents and Workflows (09:29)
  • Quiz 10.1
  • Building a CVE Finder Workflow (07:30)
  • IAM Policy Analyzer Workflow
  • Creating a Recon and CVE Analyzer Agent (08:46)
  • Quiz 10.2
  • Building Scanner Using FastAPI (06:01)
  • Agent to Agent Communication Protocols (10:33)
  • Quiz 10.3
  • Homework

11 Human Augmentation for Security Tasks

  • About this Module
  • Glossary
  • Solving Problems Using Fabric - An Introduction (22:51)
  • Installing & Configuring Fabric (17:48)
  • Quiz 11.1
  • Lab - Knowledge Mining & Distillation (15:31)
  • Lab - Breaking Language Barriers (09:14)
  • Quiz 11.2
  • Lab - Using Fabric to Parse Logs and Email Headers (21:09)
  • Lab - Threat Intelligence (26:03)
  • Lab - Video Knowledge Extraction and Validation (09:56)
  • Quiz 11.3
  • Homework

12 Attacking LLMs - Prompt Injection & Multimodal Attacks

  • About this Module
  • Glossary
  • Intro to Prompt Injection and Why it Matters (02:15)
  • Variants of AI Prompt Injection (10:31)
  • Quiz 12.1
  • Lab - Prompt Injection on a AI-Powered CV Screener (docx) (30:15)
  • Lab - Securing an AI-Powered CV Screener (12:33)
  • Lab - Setting Up Test LLM Applications on Google Colab (05:41)
  • Lab - Multimodal Attacks on a AI-Powered Meeting Summarizer (text, question-bot) (33:10)
  • Lab - Exploiting a AI-Powered Image Analyzer (image, context) (20:47)
  • Quiz 12.2
  • Homework
  • Share Your Feedback

13 Indepth look at Google's Secure AI Framework (SAIF)

  • About this Module
  • Glossary
  • Introduction to SAIF Fundamentals (02:07)
  • SAIF Pillars 1-3 Explained (13:47)
  • Quiz 13.1
  • SAIF Pillars 4-6 Explained (12:29)
  • Quiz 13.2
  • Lab: Checklist for Google SAIF Framework Pillars (01:27)
  • Introduction to Risks in SAIF (01:22)
  • SAIF Risk 1 - Data Poisoning (07:56)
  • SAIF Risk 2 - Unauthorized Training Data (07:58)
  • Quiz 13.3
  • SAIF Risk 3 - Model Source Tampering (07:41)
  • SAIF Risk 4 - Excessive Data Handling (05:15)
  • SAIF Risk 5 - Model Exfiltration (07:50)
  • Quiz 13.4
  • SAIF Risk 6 - Model Deployment Tampering (07:15)
  • SAIF Risk 7 - Denial of ML Service (07:19)
  • SAIF Risk 8 - Model Reverse Engineering (08:03)
  • Quiz 13.5
  • SAIF Risk 9 - Insecure Integrated Component (07:16)
  • SAIF Risk 10 - Prompt Injection (10:23)
  • SAIF Risk 11 - Model Evasion (09:17)
  • Quiz 13.6
  • SAIF Risk 12 - Sensitive Data Disclosure (09:30)
  • SAIF Risk 13 - Inferred Sensitive Data (09:10)
  • SAIF Risk 14 - Insecure Model Output (07:59)
  • Quiz 13.7
  • Lab: Applying SAIF Framework to AI-powered Web Application Firewall (21:24)
  • Homework

14 AI-powered Threat Modeling with STRIDE methodology

  • About this Module
  • Glossary
  • Introduction to Threat Modeling (02:07)
  • The Engineering Logic of Threat Modeling with STRIDE
  • In-depth Practical Threat Modeling, Attacks, and Mitigations using LLMs (28:10)
  • Quiz 14.1
  • Homework

15 Threat Modeling AI Applications

  • About this Module
  • Glossary
  • Rethinking AI Security Through Threat Modeling (05:04)
  • Trust Paradox and AI Attack Surface Map (04:49)
  • Introduction to Threat Modeling Frameworks for AI systems (03:07)
  • STRIDE Applied to AI Systems (04:13)
  • MITRE ATLAS and Adversary TTPs for AI (06:17)
  • Understanding OWASP MAS Threat Modeling Guide (05:26)
  • Deepdive MAESTRO's Seven-Layer Architecture (08:54)
  • A Practical Workflow for AI Threat Modeling Process (09:10)
  • Lab: Practical Threat Modeling for a AI Meeting Scribe Architecture using STRIDE Analysis (23:01)
  • Lab: Deepdive MAESTRO Analysis for AI Meeting Scribe by Layers (08:57)
  • Lab: MITRE ATLAS Adversary Techniques in Action for AI Meeting Scribe (08:51)
  • Lab: Attack Scenario Walkthrough: Corporate Espionage via Speech & Token Vault Breach via LLM (05:27)
  • Lab: Cross-Layer Threat Analysis for AI Meeting Scribe using OWASP MAS Guide (07:40)
  • Defense in Depth for AI Systems (08:38)
  • Homework

16 AI Governance and Compliance

  • About this Module
  • Glossary
  • The AI Governance Landscape and Global Regulatory Approaches (08:33)
  • NIST AI Risk Management Framework (06:28)
  • ISO/IEC 42001 AI Management System (07:04)
  • Understanding EU AI Act Risk-Based Classification (07:38)
  • Quiz 16.1
  • EU AI Act High-Risk AI Requirements (13:03)
  • EU AI Act Article 15 Security Checklist (12:28)
  • US Federal AI Policy Landscape (06:01)
  • US State AI Regulations (05:52)
  • India AI & Data Protection Landscape (06:40)
  • Pre-Deployment Security Checklist
  • Operational Security Checklist
  • Compliance Documentation Checklist
  • Quiz 16.2
  • Homework
  • Lab: Case Study of AI Hiring Assistant (08:25)

17 Exploiting and Securing MCP Servers

  • About this Module
  • Glossary
  • Debugging MCP Servers using MCP Inspector (20:47)
  • Quiz 17.1
  • Configuring Cursor & Claude for Interacting with MCP Servers (11:14)
  • Intercepting and Modifying MCP Server Traffic using Burp Suite (17:03)
  • Quiz 17.2
  • Exploiting Confused Deputy Attack via Delegation in MCP Servers (22:26)
  • Securing against Confused Deputy Attack (01:52)
  • Quiz 17.3
  • Exploiting Information Disclosure in MCP Servers (12:31)
  • Securing against Information Disclosure (02:02)
  • Quiz 17.4
  • Bruteforcing Attacks in MCP Servers (16:03)
  • Securing against Bruteforcing Attacks (02:30)
  • Quiz 17.5
  • Arbitrary file r/w in MCP Servers (18:14)
  • Homework

18 Vulnerabilities in "Vibe-Coded" / AI-generated Applications

  • About this Module
  • Glossary
  • 13 Common Vulnerabilities and Exploit Patterns in VibeCoded Applications (42:56)
  • Quiz 18.1
  • Pre-Launch Security Checklist for AI-Assisted Apps
  • Homework

19 Securing Apps using AI Gateways

  • About this Module
  • Glossary
  • AI Gateways Explained (06:10)
  • Migrating an App behind AI Gateway (08:46)
  • Adding Guardrails (05:52)
  • Analytics, Logs, and Rate Limits (02:40)
  • Quiz 19.1
  • Homework

20 Supply Chain Integrity in AI Apps

  • About this Module
  • Glossary
  • Intro to AI Supply Chain Risks (14:17)
  • Dependency Pinning (03:55)
  • Model Signing Explained (09:13)
  • Lab: Signing and Verifying a Model (06:57)
  • Common Model File Formats explained
  • Lab: Code Injection in Pytorch using Pickle (03:07)
  • Lab: Detecting a Backdoored Model (02:06)
  • Quiz 20.1
  • Homework
  • Share Your Feedback

21 CAISR Mock Exam

  • CAISR Mock Exam

22 Certified AI Security Researcher (CAISR) Exam

  • CAISR Exam Details
  • Next steps
  • Schedule Exam
  • Certificate of Completion

Certification

This course leads to the CAISR (Certified AI Security Researcher) certification delivered by 8KSec. The exam is included in the course. 


FAQ – Practical AI Security: Attacks, Defenses, and Applications (On-Demand)

How much does the course cost?
The course costs NOK 13,500 excluding VAT. Norwegian VAT of 25% applies, bringing the total price to NOK 16,875 including VAT.

How is the course delivered?
The course is delivered as self-paced, on-demand e-learning, allowing participants to study at their own pace and complete the practical exercises when convenient.

How long does the course take?
The course includes approximately 19.5 hours of video content across 136 videos. The estimated completion time is 3–5 months, depending on experience and study pace.

How long will I have access to the course?
Participants receive 365 days of access after purchase, giving them time to review the lessons, complete practical exercises and prepare for certification.

Can the course be delivered as in-company training?
Yes. In addition to individual on-demand e-learning, the course can be arranged as an in-company training programme for organisations that want to develop AI security expertise among their employees. Contact Glasspaper to discuss organisational training requirements and delivery arrangements.

Who should attend this course?
The course is designed for professionals interested in applying artificial intelligence to cybersecurity and understanding how modern AI systems can be attacked and secured.

  • Cybersecurity professionals and security engineers
  • Penetration testers and ethical hackers
  • Application security specialists
  • AI engineers and software developers
  • Security researchers
  • SOC analysts and incident responders
  • Professionals responsible for securing AI applications and infrastructure

What prerequisites are required?
Participants should have a working knowledge of cybersecurity and penetration testing fundamentals, along with an understanding of basic programming concepts in a language such as Python.

What will I learn during the course?
You will develop practical skills in both offensive and defensive AI security, including:

  • Understanding AI, machine learning and large language models
  • Building AI applications using LLMs, RAG and agent frameworks
  • Identifying and exploiting prompt injection vulnerabilities
  • Understanding security risks in Model Context Protocol (MCP) servers
  • Developing AI-powered security automation and agents
  • Performing AI threat modeling and security assessments
  • Securing AI applications using guardrails and AI gateways
  • Protecting AI software supply chains
  • Applying AI security governance and compliance frameworks

Is the course hands-on?
Yes. The course includes practical labs covering AI application development, security testing, prompt injection, MCP server security, threat modeling, model fine-tuning and defensive controls.

Which technologies and frameworks are covered?
The course introduces technologies and frameworks including:

  • Hugging Face and Transformer models
  • LangChain and LlamaIndex
  • OpenAI APIs and Anthropic Claude
  • Retrieval-Augmented Generation (RAG)
  • Model Context Protocol (MCP)
  • FAISS and vector databases
  • FastAPI and AI security automation
  • Google Secure AI Framework (SAIF)
  • STRIDE, MITRE ATLAS and MAESTRO
  • NIST AI Risk Management Framework
  • ISO/IEC 42001 and the EU AI Act
  • AI gateways, guardrails and supply chain security

Does the course cover offensive AI security?
Yes. Participants explore security testing techniques including prompt injection, jailbreaking, attacks against AI-powered applications, MCP server vulnerabilities, and AI-assisted vulnerability research through controlled lab environments.

Does the course cover defensive AI security?
Yes. Defensive topics include secure AI application design, MCP server protection, AI gateways, input and output guardrails, threat modeling, security monitoring, supply chain integrity and responsible AI governance.

What equipment do I need?
Participants need a laptop with at least 8 GB RAM, 40 GB of available disk space and administrative access to install and configure the required lab tools.

What language is the training delivered in?
The video instruction is in English. Subtitles are available in English, German, Spanish, French, Arabic and Japanese.

Does the course prepare me for certification?
Yes. The course provides a pathway to the Certified AI Security Researcher (CAISR) certification, covering foundational AI concepts, offensive security techniques and defensive AI security practices.

Is the certification exam included?
Yes. The CAISR certification exam is included in the course.

What makes this course valuable?
The course combines AI development with practical security expertise. Participants learn how to build AI-powered applications, identify vulnerabilities, implement defensive controls and apply security frameworks throughout the AI lifecycle. Its self-paced format also allows professionals to develop these skills alongside their regular responsibilities.