The Dark Side of AI: Jailbreaking, Injections, Hallucinations & more

अवलोकन

Step over to the dark side and learn about the vulnerabilities, exploits, and unintended consequences that AI models like LLMs suffer from, with hands-on prompting and exercises.What jailbreaking models involves and how to do it yourselfUnderstanding vulnerabilities inherent to models, including prompt and data leakageThe risks of exposing LLMs to proprietary or sensitive dataExploring the toxicity and bias inherently built into different modelsReal-world tests using ChatGPT, DeepSeek and other modelsExperiment with steering an LLM's neurons to prevent hallucinations

पाठ्यक्रम

•   Introduction
Welcome to The Dark Side (Intro to Guardrails and Jailbreaking)
Exercise: Meet Your Classmates and Instructor
Course Resources
•   The Dark Side of AI
Jailbreak! (The DAN Prompt)
Exercise: Create Your Own Jailbreak
Many Shot Jailbreaking
Prompt Injections - Part 1
Prompt Injections - Part 2
Thinking Like LLMs - Multi-Modal Injection
Leaking - Part 1 (Prompt Leaking)
Leaking - Part 2 (Data Leaking)
Exposure
Poisoning
Toxicity
Hallucinations
Thinking Like LLMs - Big vs Small
Challenge: Conduct Your Own Mechanistic Interpretability Research on Hallucinations
Challenge Instructions
Leaderboard: Mechanistic Interpretability
The Model Card
Model Cards Deep Dive
Exercise: Explore the Model Card for GPT-o3-mini and Learn Something New!
•   Where To Go From Here?
Let's Keep Learning Together!
Review This Byte!

---

द्वारा पढ़ाया गया

Scott Kerr

---

विषय

Computer Science