Securing Agentic AI With PyTorch: Threat Modeling & LLM Red Teaming in Practice - Valeri Milke

This talk focuses on securing agentic AI systems built with PyTorch, highlighting the unique security challenges posed by autonomous decision-making and complex reasoning. The speaker presents a security-first approach that combines AI threat modeling with practical large language model (LLM) security testing to identify risks in various aspects of agentic AI, including prompts and model interactions. By introducing MAESTRO-based AI Threat Modeling and applying the OWASP LLM Top 10 and Testing Guide, the talk offers insights into improving security within PyTorch architectures. The session also features a live demonstration of a prompt injection attack against an agentic workflow, illustrating the potential vulnerabilities and providing developers with effective strategies to detect and mitigate these risks during the AI development lifecycle. Attendees can expect to gain valuable techniques for integrating AI security testing and threat modeling into their own systems.