A robot can follow its safety rules to the letter and still behave dangerously if the information it decides on has been falsified. This is the central point of an analysis by VicOne, a cybersecurity company that protects vehicles and robots from attacks able to alter their behaviour. The problem is not the safety functions themselves, which can work correctly, but the data that feeds them: if the input lies, the protection applies the wrong rule with perfect consistency.
The distinction matters for anyone working with physical systems. An industrial robot or autonomous vehicle typically has several layers of control: sensors that measure distance, speed and the presence of people, logic that decides when to stop or slow down, and limits set by the design. All of these layers assume the incoming signals are true. A false signal, inserted at the right point, can lead the machine to consider a situation safe when it is not, without any rule being broken in the traditional sense. The system is simply deceived.
For this reason VicOne argues that safety testing for robots cannot stop at checking that protective functions activate during faults or random errors. It must also assess what happens when the information is deliberately altered. The message to teams that develop and certify these systems is to strengthen testing and produce documented evidence of that resistance, not only of normal operation.
The topic fits into a context where the autonomy of physical systems is growing fast. In recent months AlexTech has reported how AI agents already attempted real attacks during tests, how the UK's NCSC raised an alarm about post-incident detection, and how Exein raised 270 million dollars precisely to protect so-called Physical AI, meaning robots, drones and autonomous vehicles that operate in the physical world. Securing systems that act on the real world has therefore become both an investment area and a concrete concern.
The question also concerns companies that adopt these systems in factories and workplaces. When a robot performs tasks near people, its reliability depends not only on how well it recognises the world, but on how hard it is to make it believe a different reality. For firms integrating robots into their processes, the question becomes how the data the robot uses to decide is protected, alongside which rules are applied to it.
The analysis's closing recommendation is simple: treat input as an attack surface. Every sensor, communication channel or data source that influences a safety decision should be assessed as a possible vector for manipulation, and tests must show how the system reacts when that data cannot be trusted.
No comments yet. Be the first!