What Lies Beneath the Surface? Evaluating LLMs for Offensive Cyber Capabilities

This 40-minute Black Hat conference talk explores a rigorous methodology for evaluating Large Language Models' (LLMs) potential offensive cyber capabilities. Discover how current risk assessments of LLMs fall short by only testing responses to open-ended hacking challenges, creating a false sense of security. Learn about a multifaceted evaluation framework that employs prompting, simulation, and emulation on real cyber targets to measure graduated risks and determine if LLMs pose genuine offensive cyber threats to systems. The presentation includes detailed explanations of the evaluation methodology, technical implementation tools, results from initial LLM evaluations, and a live demonstration of an LLM being assessed for offensive cyber capabilities. Presented by a team of cybersecurity and AI engineers from MITRE, this talk provides valuable insights for cyber defenders seeking to understand, forecast, and prioritize defenses against potential LLM-enabled cyber threats.