Drone-Bench: Tracking simple drone surveillance capabilities of frontier models

Researchers have released Drone-Bench, a benchmark that measures how well AI models can write code to control drones for surveillance tasks, including reconstructing environments, localizing positions, navigating autonomously, detecting specific people, and following targets. The benchmark reveals that frontier AI models are rapidly advancing in physical autonomy capabilities, with the strongest models approaching or exceeding human-baseline performance on these surveillance tasks. The researchers emphasize this matters because as AI becomes increasingly capable of controlling physical systems, understanding and addressing potential misuse risks becomes critical for public discourse, not just AI labs.