RedVLA: First red teaming framework for detecting unsafe behaviors in Vision-Language-Action models before deployment

Researchers propose RedVLA, a two-stage red teaming framework designed to systematically uncover unsafe physical behaviors in VLA models (AI systems that combine vision, language, and robotic action capabilities). The framework uses Risk Scenario Synthesis to construct initial risk scenes by identifying critical interaction regions from benign trajectories, then applies Risk Amplification through gradient-free optimization to ensure stable unsafe behavior elicitation.

Experiments on six representative VLA models show RedVLA achieves an Attack Success Rate (ASR) up to 95.5% within 10 optimization iterations, uncovering diverse unsafe behaviors. The researchers also propose SimpleVLA-Guard, a lightweight safety guard built from RedVLA-generated data to mitigate these risks.

Code, data, and assets from the study are made available publicly, enabling further development of safety mechanisms for VLA model deployment.