CS Lecturer at USyd · ARC DECRA Fellow · AI Agent Security & Reliability
Supervision
A selection of students and collaborators whose work shows what it is like to do research with me.
Ziyue is my first PhD student. He formally joined the PhD program in 2026, after we began working together in 2025. Our projects include EvoHunt, a self-evolving security agent; TxRay, which reproduced 92% of real DeFi attacks with executable PoCs; a Binance bug bounty in 2026; and A2, which found 104 real-world Android zero-days in 2025.
I am Michael's secondary supervisor. We work together on finding vulnerabilities in cryptographic proofs.
I am Xiangwen's secondary supervisor. We work on AI for security and blockchain security, grounded in real-world validation. Our most significant result so far is a medium-severity Ethereum bug bounty awarded in 2025.
Shanshan works on computer-use and mobile-use agent security and benchmark reliability. His Honours thesis, AletheiaFoundry, confirmed policy violations in 70.9% of eligible attack runs across five mobile agents. This work also led to Can Agent Benchmarks Support Their Scores?.
Shane used unlearning to investigate how much security-benchmark performance comes from memorisation rather than genuine reasoning. His thesis exposed 22.5% score inflation in one security benchmark.
Strick works with me in his spare time, outside any formal degree or research program. In 2026, we co-authored EvoHunt and When Agents Overtrust Environmental Evidence.