| Description: |
Modern Large Language Models (LLMs) have shown strong capabilities in
vulnerability discovery, exploit explanation, and penetration testing
assistance. However, their tendency to hallucinate facts and produce
logically inconsistent recommendations makes them difficult to trust in
security-critical contexts. This thesis proposes the design and
implementation of a penetration testing framework where an LLM acts as
an intelligent assistant, while a formal logic engine serves as the
authoritative reasoning layer.
The framework would combine an LLM with a Prolog-based knowledge and
inference system capable of representing hosts, services,
vulnerabilities, trust relationships, attack preconditions, and
exploitation rules. Instead of allowing the LLM to directly suggest
actions, all proposals would be translated into logical assertions and
verified against the rule engine before execution. The project could
explore topics such as symbolic reasoning, explainable AI, attack graph
generation, and automated validation of penetration testing workflows.
The thesis may include the implementation of a prototype capable
of interacting with common pentesting tools such as Nmap or Metasploit,
while maintaining a formally verifiable internal model of the target
environment. An evaluation could compare the reliability and consistency
of the hybrid system against purely LLM-based approaches, particularly
in terms of hallucination resistance, reproducibility, and explainability.
|
