Shujaat Mirza

I want to understand how frontier AI can fail in ways we don’t expect, and to catch those failures before they happen. On Microsoft’s AI Red Team, I build attacks and evaluations to investigate emerging risks, with a focus on autonomy, loss of control, and multi-agent security.

I completed my PhD at NYU’s Center for Cyber Security, advised by Dr. Christina Pöpper. Previously, I was a Research Manager at MATS, overseeing projects on evaluation awareness and reward hacking. I also worked on privacy-preserving machine learning at Spotify Research.

Recent News

Oct 2026 Presented "Beyond Single-Agent Safety: Risks in Agent Networks" at the Advancing the Science of AI Agent Evaluation and Governance workshop, co-hosted by Carnegie Mellon University and Microsoft Research.

Oct 2026 Serving on the program committees for USENIX Security 2027 and AAAI 2027.

Aug 2026 Can a model's reasoning look benign while hiding a backdoor? In Evading Chain-of-Thought Monitoring Through Model Poisoning, we show how poisoned reasoning models can conceal backdoor behavior from their chain-of-thought traces.