Hi, I’m Arastoo Zibaeirad.

I am a Ph.D. student in Computer Science at UNC Charlotte, advised by Prof. Marco Vieira. I work on verification and reward design for long-horizon software agents, using vulnerability reasoning as the testbed. I build systems that read code and reason about security, and I care just as much about whether those systems can be checked when they get it wrong.

Two of my papers were accepted to the ISSRE 2026 Main Research Track. AutoTrace is an agentic pipeline that localizes vulnerability triggers by exploring a code property graph layer by layer, with every reported trigger backed by explicit graph evidence rather than ungrounded model judgment. CWE-Trace diagnoses what fine-tuning actually buys an LLM on systems software, and finds calibration without comprehension: output distributions adapt to the training data while the underlying security reasoning stays absent.

My earlier work focuses on large language models for vulnerability detection, automated program repair, and benchmark development, including VulnSage on zero-shot reasoning, VulnLLMEval on evaluation, and an ICSE 2026 workshop paper comparing diverse LLMs on detection and repair.

My current work asks what I think is the harder question. How do you check whether a model’s security claim is actually true? Training against verifiable rewards works beautifully in math and code, where an answer either passes the test or it doesn’t. Security breaks that assumption. A sanitizer gives near-perfect ground truth for bugs that crash, but says nothing about the logic and authorization flaws that never crash at all. The usual fallback, asking another model to grade the answer, covers everything but can be talked into agreeing with a confident wrong one.

I am studying a third option: program-analysis verifiers that demand a real dependency path through the code rather than a persuasive paragraph. Execution-free like a learned judge, but not something a model can argue its way past. What interests me is the tradeoff between them. What can each kind of checker actually score, and how often does each one get fooled when a model optimizes against it?

I love connecting with other researchers and practitioners. If you’re interested in collaboration, have a question about the work, or are building in this space, please reach out.