Maria-Irina Nicolae, PhD
Lead Researcher in AI and Security at Bosch Center for AI
I am an AI researcher with over 10 years of experience at the intersection of AI and security.
My work today centers on making AI systems more secure, robust, and trustworthy, with a focus on LLM evaluation and adversarial ML. Beyond securing AI, I have also contributed to the usage of AI in security, including ML-based fuzzing and intrusion detection.
Research interests
- Security for machine learning, adversarial attacks
- LLM evaluation, alignment and trust
- Machine learning for security: fuzzing, intrusion detection, AI agents
Projects
- Transformer-based CAN fuzzing – 20x increase in fault discovery for ECU testing and automotive protocols.
- Security Garten – editor of a technical blog on security, privacy and safety.
- Adversarial Robustness Toolbox (ART) – open-source library for evaluating and defending ML models.
Selected talks
Experience
- Senior Researcher — AI Security & Trust, Bosch Research, 2019 – present
- Researcher — Adversarial AI, IBM Research, 2017 – 2019
- PhD in Representation Learning, Jean Monnet University, 2013 – 2016
Publications
Google Scholar | DBLP
2025
- Z. Zhuang, **M.-I. Nicolae**, H.-P. Wang, M. Fritz. *ProxyPrompt: Securing System Prompts against Prompt Extraction Attacks.* **2025**.
[arXiv](https://arxiv.org/abs/2505.11459)
- Z. Zhuang, H.-P. Wang, **M.-I. Nicolae**, M. Fritz. *Stealix: Model Stealing via Prompt Evolution.* **ICML 2025**.
[ICML page](https://icml.cc/virtual/2025/poster/44026) | [arXiv](https://www.arxiv.org/abs/2506.05867) | [project page](https://zhixiongzh.github.io/stealix/)
- Y. Shen, Z. Zhuang, K. Yuan, **M.-I. Nicolae**, M. Fritz. *Medical Multimodal Model Stealing Attacks via Adversarial Domain Alignment.* **AAAI 2025**.
[AAAI page](https://ojs.aaai.org/index.php/AAAI/article/view/32734)| [arXiv](https://arxiv.org/abs/2502.02438) | [blog post](https://medium.com/security-garten/stealing-medical-ai-with-cat-photos-the-hidden-threat-to-radiology-models-292324b022c5)
2024
- Z. Zhuang, **M.-I. Nicolae**, M. Fritz. *Stealthy Imitation: Reward-Guided Environment-Free Policy Stealing.* **ICML 2024**.
[PDF](https://proceedings.mlr.press/v235/zhuang24a.html) | [arXiv](https://arxiv.org/abs/2405.07004) | [code](https://github.com/boschresearch/stealthy-imitation) | [project page](https://zhixiongzh.github.io/stealthy-imitation/)
2023
- **M.-I. Nicolae**, M. Eisele, A. Zeller. *Revisiting Neural Program Smoothing for Fuzzing.* **ESEC/FSE 2023.**
[PDF](https://dl.acm.org/doi/10.1145/3611643.3616308) | [arXiv](https://arxiv.org/abs/2309.16618">arXiv) | [MLFuzz repo](https://github.com/boschresearch/mlfuzz)
Patents
20+ issued US / EU / CN patents in adversarial ML, LLM security and embedded AI. Full list here.