Description:

n this thesis, you will investigate whether AI models genuinely reason about security vulnerabilities or simply adapt their answers to contextual information such as “the system was breached” or “the code passed all tests.” You will build an experimental benchmark, evaluate modern AI models on realistic vulnerable and secure code, and develop a method for making AI security reviews more evidence-based and trustworthy. The main outcome will be an open evaluation framework and a prototype AI-assisted security reviewer that supports its decisions with concrete technical evidence.