Research
152d agoNew Benchmark Exposes Agent Safety Risks
BeSafe-Bench, a new safety benchmark, exposes hidden behavioral risks of autonomous agents across web, mobile, and embodied vision-language domains. By testing agents in realistic functional environments, it sets a global standard for evaluating safety, guiding developers and regulators worldwide.
The Signal
The benchmark’s rigorous instruction space helps prevent unintended actions, fostering safer AI deployment, echoing efforts by OpenAI and Google.