The Agents of Chaos: AI Driven Malware Generation

Abstract

AI agents are rapidly making their way into offensive security, with discussions mainly revolving around AI assisted coding, the use of AI at runtime, or attacks against AI systems themselves. In this work, we explore a different idea: what happens when AI is responsible for the entire malware development process? More specifically, can local models do this too? We built an agent in which different models plan, implement, validate, and iteratively improve malware without human involvement. We evaluate different language models, prompting techniques, and workflows, comparing their impact on code quality, creativity, and evasiveness. This poster presents the architecture of the agent, the challenges encountered during development, and the techniques that solved them, and enabled quick generation of a large number of samples. We show how the agent improves randomization, helps break signature-based detection, and enables quick generation of thousands of different samples. Finally, we will share how real-world attacker use these techniques and discuss what this means for both offensive and defensive security researchers.