BraveGuard: From Open-World Threats to Safer Computer-Operating Agents

BraveGuard is a self-evolving defense framework that trains guardrail models from open-world threat signals and real agent trajectories to defend against the evolving risks faced by computer-operating agents. The framework mines emerging attack patterns from recent research, instantiates them as executable tasks and collects agent rollouts, derives trajectory-level supervision signals, and repeats the process as new threats emerge to form an adaptive defense loop. Guardrails with backbones such as Qwen3-Guard and Llama-Guard show better detection performance on trajectory-level agent safety benchmarks; on AgentHazard, compared with off-the-shelf guardrail models, accuracy under the average guard model setting rises from 38.79% to 82.38%.