Thomas Wolf introduces cybersecurity as a demanding test of dynamic reasoning rather than a simple search for code defects. Uri Rolls describes how more capable models change attack economics, while defenders still need systems that operate at scale with limited human intervention.
The presentation focuses on access-control tasks in large, connected environments. Models receive black-box tools rather than source code or internet access, and deterministic graders measure each step of a multi-stage attempt. An illustrative chain shows why gathering facts is not enough when a model misses the logical relationship between different permission checks.
Uri Rolls shows difficult evaluation results and traces of prolonged unsuccessful attempts. Thomas Wolf closes by emphasizing high-quality data, post-training and fast specialized models, arguing that open collaboration can strengthen defense rather than relying on only a few closed-model providers.
Watch on YouTube
