PhD Research, George Mason University
Advisor: Prof. Marcos Zampieri
2026 – Present
Fairfax, VA
Multilingual Robustness of Code LLM Guardrails
Under Review- Investigated whether Code LLM safety guardrails generalize beyond English to code-mixed and transliterated malicious inputs.
- Co-developed MALICE, a benchmark of approximately 250K adversarial prompts across 18 languages, derived from nine security datasets with semantic-quality filtering.
- Evaluated nine models and safety tools, including Claude 4.6 and GPT-5.2, on malicious-intent detection and code-generation refusal.
- Exposed an English-centric safety gap: the strongest classifier dropped from 0.93 to 0.21 accuracy, while leading models produced malicious code at up to three times their English rate.
Computer Science Assessment in the LLM Era
Under Review- Studied how widespread LLM use weakens the connection between submitted programming artifacts and demonstrated student competence.
- Conducted a systematic review that screened 1,714 studies and synthesized evidence from 135 peer-reviewed papers across ACM, IEEE, ACL, and Springer.
- Found that 98.5% of reviewed studies identified vulnerable assessment methods and 97.8% reported less trustworthy evidence of at least one competence.
- Developed a taxonomy of assessment redesigns and identified a recurring trilemma among authenticity, scalability, and trustworthiness to guide AI-era computing education.
AI
Safety
Code
LLMs
Multilingual
NLP
AI
Evaluation