Building an LLM Agent to empower Malware

In autonomous reinforcement learning (RL), high-level decisions are often not sufficient on their own. Once a promising direction has been identified, procedures may need to be adapted to the current context, previous outcomes, and constraints of the environment.

This thesis investigates how a large language model (LLM) agent can support context-aware adaptation of procedures (RL actions) in a Cybersecurity Offensive AI system. The goal is to study how script procedural variants can be generated or refined in a structured way so that they better match the current experimental situation. The thesis is part of a broader Cybersecurity Offensive AI research line at the intersection of intelligent agents, multi-agent systems, reinforcement learning, large language models, and controlled cyber experimentation. Strong results may contribute to a scientific publication.

Sources:

• Good Python programming skills.
• Prior coursework or experience in machine learning and NLP
• Familiarity with large language models (LLMs) workflows
• Basic understanding of reinforcement learning (RL)
• Comfort with experimentation and result interpretation