BSidesSF 2026 - Increasing the Analysis Surface of Large Language Models (Stephen Brennan, Ulrich)

This talk focuses on enhancing the analysis surface of large language models (LLMs) by introducing attention-based analysis. The speakers, Stephen Brennan and Ulrich, explain how current moderation techniques, which often rely on black box filtering, can be easily circumvented. They discuss the connection between unsafe prompts and attention patterns, demonstrating how understanding model internals can lead to improved attack detection in LLMs.