Robot Safety Systems Can Be Fooled by Manipulated Inputs, VicOne Researchers Warn

Even when a robot's safety functions are working as designed, manipulated sensor inputs could lead to unsafe behavior, according to researchers at VicOne LAB R7. In tests, a robot dog changed its movement after reading text on a poster as an instruction, and a simulated hospital service robot altered its behavior in response to crafted audio inaudible to humans. These demonstrations show how visual or audio inputs can be exploited to override a robot's intended task without changing its programming. The concern extends to independent protective systems that rely on sensor data to trigger safety stops. At a robotics bug bounty event, researchers injected a ROS 2/DDS message into a robot that was expected to remain still under a safe-control setting, yet the robot moved. Similarly, a humanoid's balance controller could be compromised if a vulnerable gyroscope is exposed to sound at its resonant frequency, causing false tilt readings and unnecessary corrections. While other safeguards might interrupt such attack chains, the key question is whether those protections use independent information or rely on the same manipulated inputs. VicOne highlights that accidental faults and deliberate cyber manipulation can lead to similar safety outcomes, from loss of control to physical harm. This challenges the assumption that safety systems will always function correctly if they are operating as specified. As robots become more integrated into workplaces and public spaces, ensuring the integrity of their sensory inputs is as critical as the safety functions themselves. The next step for robot makers and users is to audit how safety decisions depend on potentially untrusted data.