Also on Google Scholar.
arXiv preprint · 2026
Models responsible AI release as a bilevel Stackelberg game, showing that defender welfare depends on the capability gap rather than the shared capability level — so for dual-use models the decisive lever is the sequencing of access, not the deploy-or-withhold threshold.
CyCon 2026 — 18th International Conference on Cyber Conflict, NATO CCDCOE · 2026
Implements autonomous offensive red-team agents with deep multi-agent RL in the NASim environment, showing that MARL agents learn coordinated attack patterns such as lateral movement and capture the flag, while identifying training stability at scale as the primary bottleneck.
IEEE Swiss Conference on Data Science 2025 · 2025
Maps the attack surface of Retrieval Augmented Generation pipelines end to end, pairs each risk with a mitigation, and combines them with existing security standards into a framework for building trustworthy RAG systems.
ICMCIS 2025 - International Conference on Military Communication and Information Systems · 2025
A survey of multi-agent reinforcement learning for automated cyber defense, covering intruder detection, lateral movement containment, and the role of Cyber Gyms in training and validating autonomous defense agents.