The Workshop on Securing AI Research in Latin America was a 3-day workshop held in Santiago, Chile, from May 12 to 14, 2026, and organized by Deloitte in partnership with the U.S. Department of State.
Ana Carolina Erthal, Ria Deane, Juan Belieni, Gustavo Ewbank Rodrigues Danon
Project developed during Apart Research’s AI Control Hackathon. We studied whether AI control monitors remain reliable when attackers exploit unfamiliar vulnerabilities. Using the ControlArena Bash setting, we augmented a Docker environment with synthetic vulnerabilities and evaluated monitor performance under fully informed, partially informed, and uninformed knowledge conditions. We found that partial knowledge provides little improvement over complete ignorance, while full information substantially reduces attack success. This suggests that monitors may rely on explicit prior knowledge of attack surfaces rather than robust general reasoning about suspicious behavior.
Juan Belieni, Ana Carolina Erthal, Eliezer de Souza da Silva, Diego Mesquita
Machine unlearning enables the removal of specific knowledge from trained models without full retraining. While effective methods exist for single deletion requests, handling sequential requests in large language models (LLMs) remains underexplored. In this setting, we observe that gradient interference between successive unlearning steps degrades prior objectives. We propose ONPO (Orthogonal Negative Preference Optimization), which projects each step’s update onto the orthogonal complement of a low-dimensional subspace spanned by cached gradients from previous unlearning requests. This preserves prior unlearning objectives with minimal per-step overhead. On the TOFU benchmark, ONPO achieves a better trade-off between forgetting quality and model utility than existing methods.
AISB is a 4-week long intensive program to bring researchers and engineers up to speed on security fundamentals for AI systems. During this program, I developed a replication of the watermark stealing attack.
ML4Good is a bootcamp focused on AI Safety upskilling, including workshops on interpretability, alignment and governance of artificial intelligence. In this edition, I participated as a teaching assistant.
ML4Good is a bootcamp focused on AI Safety upskilling, including workshops on interpretability, alignment and governance of artificial intelligence. In this edition, I developed a Mechanistic Interpretability Course.