Blog

Matteo Boglioni

Research Intern at Mila, previously Computer Science MSc at ETH Zürich.
I’m particularly interested in red-teaming and model auditing as proactive techniques to uncover existing models’ vulnerabilities. I aim to assess and expose these flaws and inconsistencies, both in performance and security, to increase robustness and reliability towards distribution shifts, adversarial inputs, and unintended behaviors. I believe this step is crucial for a responsible and ethical deployment of AI.

Paper Accepted at ICLR 2026! Optimizing Canaries for Privacy Auditing with Metagradient Descent

I am excited to share that our paper, “Optimizing Canaries for Privacy Auditing with Metagradient Descent,” has been accepted at ICLR 2026! You can find...