ArmedAgent test finds safety risks when coding agents move robot arms

Researchers introduced ArmedAgent, a benchmark for the safety of coding agents that control an actual robot arm. These agents use a frontier model to write code and move the arm with no task-specific demonstrations or retraining. The test operates in simulation and on a physical arm. The authors tested some frontier models, with GPT-6-Astra, and found substantial safety risks. The abstract gives no rates.

Two new risks

First, the wide knowledge of the models lets agents do tasks that older robot AI could not do. Some of these tasks are dangerous. Then, agents read web pages and emails. Indirect prompt injection (hidden commands in data) can then go into the physical world.

The proof

The test includes robustness, malicious use and indirect injection. The source gives no scores, no rates of error and no ranking of models. It reports only that the results show substantial risks.

What is not known

The source does not tell which models had the lowest scores or how frequently. It does not name the agent frameworks, only that they are in wide use. The authors call the test the first systematic test of its type. That is their claim.

Insight: Readable robot code does not remove the risk of web text

Berkeley and NVIDIA show coding agents that write robot code that a person can read. A different paper reports that a usual coding task can make an agent write dangerous content. ArmedAgent adds an arm to an agent that reads web pages. Readable code reads as a method to examine the agent, but it does not stop hidden commands in data. Look for a robot vendor that tests this.