Secure Reasoning Lab

Secure Reasoning Lab

LLMs, cyber reasoning, and AI security at NYU Abu Dhabi.

Research

SRL studies how learning systems can take on security-critical work, and how to keep those systems trustworthy when they do. The themes below organize current and recent projects.

Cyber Reasoning Systems

We build agents and tools that find bugs and produce patches in real software.

  • PatchAgent — a practical program-repair agent that mimics human expertise. USENIX Security 2025 (Long Talk); CSAW 2025 Technical Impact Runner-up. This line of work also supported Team 42-b3yond-bug in the DARPA AI Cyber Challenge.
  • BandFuzz — collaborative fuzzing with reinforcement learning. First prize at the SBFT 2024 Fuzzing Competition; later used as a fuzzing component in AIxCC.
  • Locus — agentic predicate reasoning for directed fuzzing. ICSE 2026.
  • NSOC — the NYU Software Supply Chain Security Operations Center, co-founded with Justin Cappos. Students serve as year-long liaisons to open-source ecosystems and apply AI tools to vulnerability discovery and patching. Announcement.

Secure and Robust LLMs

We evaluate and harden language models against jailbreaks, prompt injection, and brittle alignment.

  • LLM-Fuzzer / GPTFuzzer — automated jailbreak assessment. USENIX Security 2024; Geekcon 2023 Breakthrough Award. The tool has been adopted by major LLM providers and integrated into Microsoft Azure PyRIT.
  • PromptFuzz — fuzzing for prompt-injection testing. IEEE TIFS 2026.
  • Mind the Inconspicuous — hidden weaknesses in aligned LLMs’ ethical boundaries. USENIX Security 2025 (Long Talk).
  • Custom GPT prompt injection — risk assessment of 200+ custom GPTs, covered by WIRED.

Reasoning and Explainability

We try to make model decisions more effective and more inspectable, especially on long-horizon tasks.

  • GPO — learning from critical steps to improve LLM reasoning. NeurIPS 2025; covered by MIT Technology Review China.
  • EntroPO — entropy-enhanced multi-turn preference optimization for coding agents. 1st on SWE-bench Lite and 5th on SWE-bench Verified among open-weight models.
  • RICE — explanation-guided reinforcement learning. ICML 2024 Spotlight.
  • StateMask and AIRS — step-level explanations for deep RL, including security applications. NeurIPS 2023; USENIX Security 2023.

Blockchain Security

  • BlockScan — detecting anomalies in blockchain transactions. NeurIPS 2025.