Manipulated Sensor Data Exposes New Vulnerability in Robot Safety Systems

Researchers at VicOne LAB R7 demonstrate that robots can follow safety protocols while still acting dangerously if their sensor inputs are deliberately manipulated or corrupted. Tests show that adversarial visual patches, crafted text, and inaudible audio can change robot behavior without triggering safety mechanisms. The findings highlight a critical gap where protective systems designed to prevent harm may themselves rely on compromised information sources.

Researchers at VicOne LAB R7 documented multiple attack vectors where robots misinterpret their environment due to manipulated inputs. Tests included adversarial visual elements in camera feeds, specially crafted text treated as commands, and inaudible audio signals—all capable of altering robot behavior without triggering intended safety responses. A particularly concerning finding involved injecting false messages into a robot's control system at a bug bounty event, causing movement that should have been prevented by safety protocols.

The core vulnerability lies in a cascading dependency: safety mechanisms designed to protect people often rely on the same sensor data that attackers can corrupt. If a robot's distance sensor reads false information, both the primary task system and the backup safety system receive identical falsehoods. This means redundancy between systems may provide less protection than assumed, especially when a single compromised input source affects multiple layers of oversight.