This research identifies significant security vulnerabilities in robotic systems that use Vision-Language-Action (VLA) models to interpret visual and language inputs.

• VLA models introduce new attack surfaces that can be exploited through adversarial inputs
• Attackers can manipulate either visual or language components to compromise robot behavior
• Both targeted and untargeted attacks were found effective against current systems
• Results highlight the urgent need for robust defense mechanisms before widespread deployment

For security professionals, this research serves as a critical warning that AI-powered robotic systems require comprehensive security testing and hardening before real-world implementation.

Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics