ArmedAgent benchmark finds safety risks in coding agents that control robot arms

Yiming Sun and Mi Zhang have introduced ArmedAgent, a benchmark for the safety of coding agents that control a real robot arm without task-specific demonstrations or retraining, in a technical report posted on arXiv.

Such agents, powered by frontier models, can do tasks earlier vision-language-action approaches could not, which opens the door to misuse, the authors write. They can also read web pages and emails, so threats there, such as indirect prompt injection, can reach the physical world through the robot.

The benchmark has 36 tasks in five scenarios covering robustness, malicious use and indirect injection, run in simulation and checked on a physical robot arm. Testing several frontier models, including GPT-6 Astra, in widely used agent frameworks, the authors found substantial safety risks in current systems.