Your Robot’s Safety Functions Already Work. What If the Input Lies?
IEEE Spectrum VicOne
Robot safety can fail if the input is lying. VicOne says that turns a “safe” system into a repeatable attack, not just a random glitch.
Based on reporting by IEEE Spectrum, VicOne — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
A robot can be doing exactly what it was built to do and still put people at risk. If the hallway looks empty when it isn’t, if a worker appears farther away than they are, or if a stop signal never lands, the machine may react with perfect obedience to bad information.
That is the point VicOne LAB R7 is pushing: safety functions do not live apart from cybersecurity. They rely on readings, estimates, messages, and other inputs. If those inputs are manipulated, the protective layer can make the wrong call just as faithfully as it would make the right one.
The article points to several examples. In published research on vision-language-action models, a patch in the camera view reduced task success in simulation. FreezeVLA showed that an adversarial image could make tested models ignore later instructions. VicOne also cites a robot-dog test using Gemma 4 E4B, where text on a poster was treated as an instruction and changed movement, plus a hospital-service-robot simulation using Nemotron on NVIDIA Jetson AGX Orin, where crafted audio inaudible to people altered simulated behavior.
The same problem shows up on the safety side. At a robotics bug bounty event, VicOne researchers injected a ROS 2/DDS message into a robot that was expected to remain still under a safe-control setting, and it moved. The article also describes a humanoid balance controller that could be misled if a vulnerable gyroscope is distorted by sound at its resonant frequency. Researchers have demonstrated that acoustic attack against drones with vulnerable gyroscopes, though not that exact humanoid failure.
VicOne’s larger argument is that deliberate attacks change how risk should be assessed. Accidental faults and hostile manipulation can lead to similar outcomes, but an attacker can time the trigger and repeat it. That means redundant sensors and independent checks deserve adversarial testing, not a pat on the head for agreeing with each other. The company also ties this to current and upcoming requirements, including China’s GB/T 45502-2025, IEC TS 63074, and the EU Machinery Regulation from January 20, 2027.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous truth robotics keeps trying to outrun: a safe system that trusts its inputs is only as safe as the easiest lie. The industry loves talking about autonomy, then gets surprised when a poster, a message, or a sensor glitch becomes the boss. Open systems make that easier to test; closed ones make it easier to pretend the problem isn’t there.
Read more about this at: IEEE Spectrum
