Julius Graf named Berkeley’s 2026-27 Cal Cybersecurity Research Fellow

UC Berkeley has named Julius Graf its 2026-27 Cal Cybersecurity Research Fellow, supporting research into security threats involving generative AI and autonomous agent systems. Graf’s project will examine how a single prompt injection could spread through interconnected AI agents and develop methods to contain such attacks.
- Julius Graf is UC Berkeley’s 2026-27 Cal Cybersecurity Research Fellow.
- The fellowship focuses on cybersecurity threats involving generative AI and agentic systems.
- Graf will study cascading attacks caused by prompt injection across AI agents.
- The project proposes mathematical models and a “cognitive firewall” for containment.
Research fellowship focuses on AI security
The fellowship is presented by the Berkeley Center for Long-Term Cybersecurity, or CLTC, with support from Cal alumnus and CLTC External Advisor Tim M. Mather, Class of ’81. CLTC has presented the Cal Cybersecurity Research Fellowship since 2019.
The 2026-27 fellowship is dedicated to research at the intersection of cybersecurity and artificial intelligence. Its focus is on emerging threats posed by and to generative AI and agentic systems, which can read, analyze, and act on outputs from other systems.
Graf’s work spans cyber risk and AI safety
Graf is a second-year Ph.D. student in Industrial Engineering and Operations Research at UC Berkeley, where he is advised by Professor Thibaut Mastrolia. His research applies stochastic control, Hawkes processes, and reinforcement learning to market making, cyber risk, and AI safety.
He holds an M.Eng. in Industrial Engineering and Operations Research from UC Berkeley and a Diplôme d’Ingénieur from CentraleSupélec. Graf is also a Graduate Student Instructor for INDENG 241, Risk Modeling, Simulation, and Data Analysis, and co-founder and CTO of EquitX, an intent-based market infrastructure for live event tickets.
Mastrolia has been an Assistant Professor of Industrial Engineering and Operations Research at UC Berkeley since July 2021. His research includes stochastic analysis and control, contract theory and mean-field games, financial engineering, cyber risk management, and AI safety.
Project targets cascading attacks between AI agents
Graf’s research abstract describes a security problem that emerges when autonomous systems treat outputs from other agents as trusted instructions. A successful prompt injection could cause a compromised agent to generate malicious emails, alter shared documents, poison memory entries, or send forged messages. The resulting borrowed authority and harvested credentials could make later breaches faster and easier to execute.
Working with Mastrolia, the Center for AI Risk Management & Alignment, and Stoa, Graf plans to model these breaches as self-exciting chain reactions using an amplified Hawkes process. The model is intended to identify operational thresholds at which attack cascades remain localized, require active intervention, or lead to network failure.
The project also proposes a “cognitive firewall” with automated threat alarms, real-time repair protocols, and circuit-breaking mechanisms. Graf’s team will calibrate and stress-test the model and firewall on a multi-agent testbed using realistic and adaptive attack scenarios. According to CLTC, the work aims to provide quantitative benchmarks for how quickly AI operators must detect, repair, and halt breaches in interconnected systems.
Berkeley CLTC: Announcing the 2026-2027 Cal Cybersecurity Research Fellow, Julius Graf


