Prompt Engineering for Security Agents - A Measurable Approach with GEPA
GEPA (Genetic-Pareto) is an LLM-based prompt optimization framework that enables measurable refinement of agent instructions without retraining. The article walks through applying GEPA to improve a CTF-solving agent by defining explicit reward functions (challenge solved, turn efficiency, duration) and providing actionable side information from test runs to guide prompt iterations. GEPA works with closed-source models via inference-only calls, requires modest compute, and offers a practical alternative to reinforcement learning for security professionals building and optimizing LLM-based agents and tools.
This post is licensed under CC BY 4.0 by the author.