ArmedAgent benchmark evaluates the safety risks of coding agents controlling robotic arms

1. 23:02
   arXiv: Jailbreaks, Prompt Injection, Poisoning & Defenses

   When coding agents are equipped with robotic arms: safety risks crossing digital and physical boundaries

   The researchers propose the ArmedAgent benchmark, which systematically evaluates the safety risks of coding agents controlling real robotic arms zero-shot and end-to-end, finding significant safety hazards in current systems across multiple frontier models. The authors state this is the first systematic benchmark for this problem. The benchmark covers three risk categories—robustness, malicious use and indirect injection—with 5 scenarios and 36 tasks in total, evaluated in simulation and validated on a physical robotic arm, with some tasks specifically targeting harmful behaviors that earlier VLA paradigms could not accomplish. The paper points out two new challenges: the broad knowledge of frontier models lets agents complete tasks that earlier VLA and WAM paradigms could not, creating new space for capability misuse; and coding agents can access digital resources such as web pages and email, so threats such as indirect prompt injection therein enter the physical world along with robotic arm control. Testing is conducted under commonly used agent frameworks and covers multiple frontier models including GPT-6-Astra.

Current attention 35·Peak in the comparable period 38 (Nov 10 22:00)·Change over the comparable past 24 hours -1%