Attacking and Defending LLMs
This workshop gives you hands-on experience with attacking large language models (LLMs) using a range of prompt-based strategies. You will actively explore how these attacks work in practice and what their impact is on real systems. The workshop also gives you insight into defensive techniques, and shows how architectural choices, testing approaches, and security observability can be used to strengthen applications built with generative models.
Content overview
- AI security fundamentals in the context of LLMs
- Prompt injection attacks and their practical impact
- Document-assisted generation and associated risks
- Guardrails and defensive techniques
- Prompt routing strategies
- Security observability for LLM-based systems
Content level
Introductory
Target audience
Developers, engineers, and practitioners who are building or planning to build systems that leverage LLMs.
Prerequisites
Basic Python knowledge and familiarity with running code locally. Some exposure to LLM APIs or AI tooling is helpful but not required.
Technical requirements
Laptop capable of running Python and local LLM tooling. Participants will receive a repository with exercises and instructions for running models locally.
Katharine Jarmul focuses her work and research on privacy and security in data science, deep learning and AI. She is author of the well received O'Reilly book Practical Data Privacy (O'Reilly 2023) and has more than 10 years experience in machine learning/AI where she has helped build large scale AI systems with privacy and security built in.