PyRIT: From LLM Security Research to Practical Attacks

No ratings

Presented at Bluehat / BlueHat India 2024 by

Richard Lundeen

The world of LLM security is wild. Papers are published at a rate where it’s hard to keep up. And because of the probabilistic nature of LLMs, these attacks can often be combined and applied in unpredictable and interesting ways. PyRIT is an open-source library developed by the Microsoft Al Red Team, and it aims to tackle this problem. PyRIT is built for researchers, pentesters, and defenders to help assess the robustness of LLM applications against different attacks and harm categories. The main scenario this talk focuses on is how PyRIt can be used to take research findings, combine them, and actually use them against a broad range of LLM endpoints. We want PyRIT to be the go-to tool researchers use to make proof-of-concept attacks and adapt them to real-world scenarios. This talk walks through some of the most impactful public research we've incorporated into PyRIT - and how you can too! We'll show patterns of how we included external research like “tree of attack”, “PAIR”, “Crescendo”, and “persuasive adversarial prompts”. Using PyRIT for testing LLM security goes beyond downloading Hugging Face models and checking their security. We'll talk about how these proof-of-concepts can be included so researchers can test their entire pipelines against these attacks Cincluding their derivations).