Automated Vulnerability Discovery and Mitigation in the AI Era

ABSTRACT

Finding security vulnerabilities in software is an important but challenging task. Traditional security testing often requires developers or security experts to manually understand a program, design test cases, and analyze unexpected behaviours. Recent AI agents and large language models provide a new possibility: can we teach AI agents to automatically understand software and help discover security vulnerabilities?

This project explores how AI agents can be used for automated software security testing. Students will work with modern AI coding agents and security testing tools, and investigate how AI can help with tasks such as understanding program interfaces, generating test programs, running fuzzing campaigns, improving test coverage, and analyzing crashes. Students will start from practical and manageable tasks, learn the necessary background step by step, and gradually explore open research problems.

Possible project topics include:

• Using AI agents to automatically test the security of open-source software.
• Using AI to analyze program crashes and security warnings.
• Building simple AI-agent workflows for software security testing.
  Through this project, students will gain hands-on experience with real-world software systems, AI agents, fuzzing, and security testing. They will also learn how research ideas are developed, implemented, evaluated, and presented.

SKILLS & EXPERIENCE REQUIRED

• Disciplines in Computer Science or Data Science
• Familiar with at least one programming language: Python, Java, C, C++, JavaScript, etc.
• Interested in AI and software security
• Passion and commitment for influential research