I am an Offensive Security Engineer at NVIDIA, working on the DriveOS platform within the automotive org. My work focuses on automating security workflows and supporting offensive security efforts, with a particular interest in leveraging LLMs to scale and augment security analysis.
More broadly, I am interested in systems security and in understanding the layers of computation that lie beneath immediate interfaces—spanning low-level components such as hypervisors, operating systems, and firmware. During my research at Univeristy of California, Santa Barbara under the supervision of Giovannia Vigna and Christopher Kruegel, I worked on operating system security, with a focus on kernel fuzzing. My broader interests include binary analysis, root cause analysis, ML applications in security, as well as ARM TrustZone, rehosting, and OS security.
If you are interested in related topics, feel free to reach out—I am always happy to connect.
Education
University of California, Santa Barbara
PhD program in Computer Science, exited with Masters
Friedrich-Alexander University Erlangen-Nuremberg
Bachelor of Science in Computer Science, graduated with distinction
Publications
ACTOR: Action-Guided Kernel Fuzzing (USENIX ‘23) pdf cite
Marius Fleischer, Dipanjan Das, Priyanka Bose, Weiheng Bai, Kangjie Lu, Mathias Payer, Christopher Kruegel, Giovanni Vigna
Syzgrapher: Resource-Centric Graph-Based Kernel Fuzzing (RAID ‘25) pdf cite
Marius Fleischer, Harrison Green, Ilya Grishchenko, Christopher Kruegel, Giovanni Vigna
Orion: Fuzzing Workflow Automation (Arxiv ‘25) pdf cite
Max Bazalii, Marius Fleischer
Projects
Speech Style Transfer
- Developed one-click pipeline for speech style transfer based on SOTA models AutoPST, AutoVC
- Evaluated the model performance on a multi-accent name dataset, live data during a demo presentation
Fake News Detection
- Developed an end-to-end data collection pipeline using the Twitter API v1 and v2 that generates a dataset for Graph Neural Network (GNN) models starting from a set of root tweets
- Evaluated the performance of existing GNN models (developed in PyTorch) on multiple social media datasets