Injection Attacks & Security in LLMs - Overview and Open Problems - 1/2 скачать в хорошем качестве

Injection Attacks & Security in LLMs - Overview and Open Problems - 1/2

Speaker: Ana Clara Zoppi Serpa (PhD Candidate - Unicamp - ex Google, Microsoft, and Amazon) Summary: In this talk, Ana delivers an overview of LLM injection attacks and discusses theopen challenges in the research community. She maps established taxonomies from recentliterature, clarifying distinctions between key categories: jailbreaks, prompt injection,fingerprinting, evolutionary algorithm-based attacks, and prompt engineering methods,highlighting notable results from the latest research. Time permitting, she will demonstrate alive attack against a target LLM. The talk concludes with open problems actively shaping the field. Bio: Ana Clara Zoppi Serpa is a PhD student at UNICAMP researching prompt injectionattacks in LLMs, building on her Master's in Computer Science and award-winningcryptography research—Best Short Paper at the Brazilian Symposium on Security in 2019.With 3+ years as a Software Engineer at Google (AI/LLM data pipelines), Microsoft (Azureinfrastructure), and Amazon (backend systems), she bridges theoretical expertise withproduction-scale implementation.