Tag
BraveGuard is a self-evolving defense framework that trains guard models using open-world threat signals and realistic agent trajectories to improve safety detection in computer-use agents, achieving significant accuracy gains on the AgentHazard benchmark.