In this talk, SANS Faculty Fellow and course author, Seth Misenar will explore potential security implications of extending LLMs beyond their initial self--supervised pre-training. Moving beyond a model’s innate limits can prove advantageous, but some approaches can also materially change the attendant risk profile. Surveying security implications of various extension strategies such as In-Context Learning, (ICL) Retrieval Augmented Generation (RAG), LLM Ensembling, Agentic LLMs, or Fine-Tuning can allow organizations to better navigate the security landscape of extending LLMs.

SANS AI Cybersecurity Forum 2024
Go-Go Gadget Cyber: Extending LLMs
Speaker: Seth Misenar

View upcoming Summits: http://www.sans.org/u/DuS