AI Security Fundamentals: The OWASP Top 10 for LLMs
For security engineers, AppSec teams and developers who need to find, explain and fix the security risks that LLM features add to real applications.
Level
Intermediate
Study time
14 hours
Lessons
18 in 6 modules
Exam
20 questions
About this course
Adding a language model to an application changes its attack surface in ways that classic AppSec checklists miss: instructions and data share one channel, retrieved documents and tool results can carry commands, the model's output is untrusted input to everything downstream, and an agent can turn a clever sentence into a database write. This course teaches you to reason about those risks precisely and to fix them with architecture rather than hope.
You start with why LLMs change the attack surface and how to draw trust boundaries for chat, RAG and agent designs. You then work through all ten risks in the 2026 OWASP Top 10 for LLM Applications (released August 2026), with a crosswalk to the 2025 numbering that many tools and reports still use: prompt injection, sensitive information disclosure, excessive agency, supply chain, data and model poisoning, unbounded consumption, misinformation, hidden context exposure (formerly system prompt leakage), vector and embedding weaknesses, and improper output handling. Each lesson pairs hypothetical scenarios with mitigations you can implement. The final module maps the risks to MITRE ATLAS techniques and the NIST AI 100-2 E2025 adversarial machine learning taxonomy, introduces adversarial testing, and gives you secure design patterns and a reusable security review checklist for any LLM feature. Framework facts are current as of October 2026.
What you will be able to do
Explain how LLM applications change the attack surface and draw trust boundaries for chat, RAG and agent architectures
Describe each risk in the 2026 OWASP Top 10 for LLM Applications and map it to its 2025 equivalent
Distinguish direct, indirect and multimodal prompt injection and choose architectural controls that limit their impact
Implement defenses for sensitive data exposure, hidden context exposure, and vector store authorization in RAG systems
Constrain agent tools with least privilege, user-scoped authorization, approval gates and budgets
Treat model output as untrusted input and apply context-appropriate encoding, validation and sandboxing
Map LLM application risks to MITRE ATLAS techniques and NIST AI 100-2 attacker objectives and capabilities
Run a structured security review of an LLM feature using a checklist, threat model and adversarial test suite
Syllabus
6 modules · 18 lessons · about 14 hours
1
Module 1: The LLM Attack Surface
Why language models change application security, where the trust boundaries are, and how the 2026 OWASP list is organized.3 lessons
Why AI Changes the Attack Surface
Anatomy of an LLM Application: Trust Boundaries and Data Flows
The OWASP Top 10 for LLM Applications 2026: Structure, Changes and How to Use It
2
Module 2: Injection and Hidden Context
Direct, indirect and multimodal prompt injection, the architectural defenses that actually limit it, and exposure of hidden context.3 lessons
Prompt Injection (LLM01): Direct, Indirect and Multimodal
Defending Against Prompt Injection: Architecture Over Filters
Hidden Context Exposure (LLM08): From System Prompt Leakage to Agent Context
3
Module 3: Data Exposure and Integrity
Sensitive information disclosure, weaknesses in vector stores and embeddings, and poisoning of data and models.3 lessons
Sensitive Information Disclosure (LLM02)
Vector and Embedding Weaknesses (LLM09)
Data and Model Poisoning (LLM05)
4
Module 4: Agency, Output and Supply Chain
Agents with too much power, model output flowing into dangerous sinks, and the models, datasets and tools you did not build.3 lessons
Excessive Agency (LLM03)
Improper Output Handling (LLM10)
Supply Chain (LLM04): Models, Datasets, Adapters and Tools
5
Module 5: Misinformation, Consumption and Adversarial Testing
Plausible but wrong output, runaway cost and denial of service, and how to test an LLM application adversarially.3 lessons
Misinformation (LLM07)
Unbounded Consumption (LLM06)
Adversarial Testing for LLM Applications
6
Module 6: Frameworks, Secure Design and Review
Mapping risks to MITRE ATLAS and NIST AI 100-2, reusable secure design patterns, and a security review checklist for LLM features.3 lessons
Mapping Risks to MITRE ATLAS and NIST AI 100-2
Secure Design Patterns for LLM Features
The LLM Feature Security Review Checklist
Exam and certificate
The exam unlocks once you have completed every lesson. It draws 20 questions from a larger bank, spread across the modules, with the answer options shuffled for each attempt. You have 45 minutes and need 80% to pass. After a failed attempt you can try again after 24 hours, up to 3 attempts in total. Feedback is given per module, never as the answer key.
Passing earns a certificate of completion and assessment from the Wirewalk Institute of AI with a unique ID that anyone can check at institute.wirewalk.com/verify. It is not an accredited qualification or a vendor certification. How certification works.
The Wirewalk Institute of AI is independent. This course teaches publicly documented products and is not affiliated with or endorsed by OWASP, MITRE ATLAS, NIST, NVIDIA garak, Microsoft PyRIT, or Promptfoo. Product and company names are trademarks of their respective owners.
Follow one fictional US distributor through four everyday AI mistakes (a guarantor's credit file uploaded to a personal account, a customer letter promising a fee waiver that does not exist, a vendor email with hidden instructions for the assistant, and a note-taker that sent a customer the internal margin floor) and practice the decisions that prevent or contain each one in three short Decision Labs.
A plain-language course on handling data safely with AI tools: what happens to what you type or upload, how to classify before you paste, how connected apps and agents change the picture, and how to report a mistake quickly.
A deep course on prompt injection: why it exists, how direct, indirect and multimodal payloads turn into data exfiltration and goal hijack in agents, why filters alone fail against adaptive attackers, and which architectural patterns and tests actually reduce risk.