Your Robot’s Safety Functions Already Work. What If the Input Lies?

VicOne LAB R7, the research arm of the Tokyo-headquartered Physical AI cybersecurity company VicOne, has published a sponsored IEEE Spectrum article laying out how manipulated inputs can produce unsafe robot behavior even when the robot's safety logic is working as designed. The central argument: safety functions have inputs of their own, and if those inputs are falsified, a protective system can make the wrong call while appearing healthy. Alongside the article, VicOne is offering a white paper titled "Why Robot Safety Depends on Cybersecurity: Protecting the Information That Safety Decisions Rely On."
The scenarios are ordinary. A mobile robot slows for a person in a hallway, a collaborative arm eases its motion as a worker approaches, a humanoid pauses to let someone pass. Each depends on a distance reading, a position estimate or a stop signal. Reverse any one of those and the robot still executes its rules, just against a false picture: the hallway reads clear with a person still in it, the co-worker appears farther away than they are, the stop signal never arrives.
Prior published research supports the general mechanism, according to the article. A study of vision-language-action (VLA) models, which map camera images and language instructions to robot actions, found that a patch placed in the camera's view reduced task success in simulated robot tests. A separate piece of work, FreezeVLA, found an adversarial image could cause tested models to ignore later instructions. The two used different methods but both show visual input interfering with an intended task.
VicOne's own tests are the most concrete part of the piece. In one, a robot dog running Gemma 4 E4B treated text on a poster as an instruction and changed its movement. In another, a hospital-service-robot simulation running Nemotron on an NVIDIA Jetson AGX Orin had its simulated behavior altered by crafted audio inaudible to people. In both, the article states, the assigned tasks did not change; the inputs did. That distinction matters for anyone building guardrails: a robot can pass a task-level check while acting on something an attacker planted.
The sharper point concerns independent protection. A protective layer may still halt a dangerous action, but it too consumes readings, estimates and messages. At a robotics bug bounty event, VicOne LAB R7 researchers injected a ROS 2/DDS message into a robot that organizers expected to stay still under a safe-control setting, and the robot moved. ROS 2 is the widely used open-source Robot Operating System middleware; DDS is the publish-subscribe transport underneath it, and message injection along that path is a recognized attack surface in robotics deployments.
The article also sketches a chain in which a humanoid's balance controller relies on a vulnerable gyroscope, and sound at the sensor's resonant frequency distorts its reported rotation, so the controller corrects for a tilt that never happened. Researchers demonstrated the underlying acoustic attack against drones with vulnerable gyroscopes; the article explicitly notes the same chain was not shown causing a humanoid to fall, and that other protections could interrupt either chain. The real question it poses for safety and security teams is whether those other protections detect the hazard independently or lean on the same manipulated data.
Deliberate attack changes the probability math behind safety assessments. A hazard analysis may dismiss a particular combination of accidental faults as vanishingly unlikely. An attacker, by contrast, can pick the moment to introduce a false input and repeat the same trigger. Harm is not inevitable, but the risk model needs revisiting. Redundancy gets the same treatment: if a single action can fool two sensors at once, their agreement reassures less than it seems. In autonomous driving research, a crafted physical object misled a system fusing camera and LiDAR perception. That study does not establish a weakness in any particular robot, the article says, but it shows why teams should test whether one manipulation can bend several checks simultaneously.
On the compliance side, the article points to three frameworks with different scopes. China's GB/T 45502-2025 addresses information security for service robots. IEC TS 63074 examines security threats that could affect safety-related control systems. The EU Machinery Regulation applies from January 20, 2027 and includes requirements to protect safety-relevant systems and data against corruption. None of these substitutes for evidence about a specific robot's design and operating conditions.
VicOne's proposed practice runs across the lifecycle: before deployment, compare normal and adversarial scenarios against the same safety limits for speed, protective distances and permitted operating areas; during operation, monitor for changes to software, models, sensor signals or behavior that could invalidate earlier results. Responses to suspicious behavior should follow policies the safety team has already approved. The company markets two tools for this, Radeis for pre-deployment validation and Rthena for in-operation visibility into risks and behavior, and runs a Robotic Hacking Community through LAB R7.
The article closes with five questions it suggests safety and security teams answer jointly: what each protective function reads, including which safeguards run independently of the robot's AI decisions; how that information is produced, transmitted and authenticated, and how missing or implausible values are handled; what happens when an input is manipulated, tested with false distances and drifting position estimates; whether one action can mislead several inputs through sensor fusion or a shared point of failure; and when evidence should be re-examined after changes to software, models, sensors or operating conditions.
The commercial framing is worth noting when weighing the findings. VicOne sells testing and monitoring software for robots and vehicles, so its conclusion that adversarial evidence belongs in safety cases aligns with its products. The individual demonstrations remain useful as illustrations: prompt-style text injection into a VLA model, ultrasound against a gyroscope, and middleware message injection are all well-documented classes of attack. My reading is that the harder engineering problem is architectural rather than detection-based, since an agent that reasons over camera and audio input necessarily blurs the line between data and instruction.
Why it matters: Teams building robots that work near people now face a safety argument that must hold under adversarial input, not only component failure, and that likely means new tests, input authentication and failure detection around ROS 2/DDS paths and fusion pipelines. As inference on devices like Jetson-class hardware becomes standard, my inference is that security review will shift toward the model and middleware layers where instruction-following and message trust are decided. The EU deadline in January 2027 gives European deployments a concrete date to work toward.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
This article is brought to you by VicOne . A mobile robot slows for someone in a hallway. A collaborative arm eases its movement as a worker approaches. A humanoid pauses to let a person pass. Each response depends on information about the robot or its surroundings: a distance reading, a position estimate, or a stop signal. But what happens when one of those inputs is wrong, and the robot still responds exactly as designed? The mobile robot reads the hallway as clear while someone is still there. The arm estimates