
Explore the LLM security landscape by examining the top 10 vulnerabilities across training, prompting, and deployment, with real-world incidents and practical mitigations.
Explore the OWASP top 10 LLM vulnerabilities, starting with prompt injection—direct and indirect forms—along with sensitive information disclosure, supply chain vulnerabilities, data and model poisoning, and improper output handling.
Explore the OWASP top 10 LLM vulnerabilities, including agency and system prompts leakage, vector weaknesses, retrieval augmented generation risks, hallucinations, misinformation, and unbounded consumption.
Explore the LLM development lifecycle—from training to deployment—identify vulnerabilities in each phase, including training data poisoning, prompting phase risks like prompt injection, and deployment API gaps, with actionable mitigations.
Explore data and model poisoning across training data, open-source model contamination, embedding and interaction attacks, and remote code execution risks, with defenses like trusted data sources, ML-BOM, red-teaming, and monitoring.
Explore how vector and embedding weaknesses in LLM systems expose data through embedding inversion, RAG poisoning, and cross-tenant leakage, and learn mitigations through strict access control and data redaction.
Explore supply chain vulnerabilities across models, data, libraries, and frameworks in the LLM lifecycle. Learn mitigations like BOMs, audits, hashes, licenses, monitoring, and patching to prevent poisoning and breaches.
Examine direct and indirect prompt injection as a security vulnerability that can bypass safety prompts and leak sensitive data. Learn defenses like input validation, regex checks, and human-in-the-loop.
Learn how large language models can disclose sensitive information like PII, financial and health data, and how to mitigate risks through output validation, data masking, and stakeholder education.
Explore improper output handling, including client-side XSS and CSRF and server-side SSRF and remote code execution, and how validation, sanitization, encoding, and parameterized queries mitigate these risks.
System prompts shape model behavior, but leakage can expose secrets, internal logic, and user roles. This enables privilege escalation and unauthorized access, prompting defenses like least privilege and real-time monitoring.
Identify misinformation in gen ai, including hallucination, bias, and overgeneralization, and its impact on healthcare and legal. Explore mitigations: rag with external databases, fine-tuning, cross-verification, human oversight, and automated fact-checkers.
Explore excessive agency in AI agents, limit permissions and plugins, enforce least privilege, and apply human-in-the-loop and input-output sanitization to prevent data leakage and unauthorized actions.
Identify unbound consumption attacks that exhaust resources through unlimited requests and resource-intensive prompts. Implement defenses with timeouts, throttling, strict input validation, rate limits, and glitch token filtering.
Reflect on what could be hardened in your AI system to make the world a safer place, and share comments and ratings to improve future teaching experience.
This course is intended to provide CPE hours; recognition of CPE credit is subject to your certification body’s approval.
Generative AI is transforming industries, but it also introduces new security risks that many organizations underestimate until a real incident occurs. This course, Gen AI Security Foundations, provides a practical and structured introduction to the most pressing security challenges that arise when working with Large Language Models (LLMs) and generative AI systems.
Across a series of focused lectures, participants will gain a comprehensive understanding of the OWASP Top 10 LLM Vulnerabilities for 2025, including threats such as prompt injection, model poisoning, sensitive data disclosure, improper output handling, excessive agency, vector database weaknesses, hallucination-induced misinformation, and unbound consumption attacks. Each vulnerability is explored through its technical background, real-world case studies, potential impacts, and proven mitigation strategies.
The training also maps vulnerabilities to the LLM development lifecycle—Training, Prompting, and Deployment—illustrating how risks emerge at different stages. Most importantly, the course emphasizes mitigation strategies. You will learn how to apply security best practices such as dataset validation, input-output sanitization, access controls, monitoring, and human-in-the-loop safeguards to reduce vulnerabilities in your AI systems.
By the end, you will be able to recognize, classify, and mitigate key LLM security risks while applying proven defense techniques to strengthen your AI solutions.
Whether you are a developer, architect, or security professional, this course equips you with the awareness and skills to harden AI systems and ensure safer, more trustworthy deployments in production environments.