For decades, robot safety has revolved around a single question: can a machine remain safe when something goes wrong? That framing assumes failures are accidental — a broken sensor, a software bug, an unexpected collision. The goal is to contain harm when normal operation breaks down.
An article in IEEE Spectrum argues that physical AI — robots that perceive and act through learned models — demands a harder question. Can a machine remain safe when an attacker changes what it sees, decides, or does, even when nothing appears wrong? The threat is no longer just a fault but deliberate manipulation, and it can be invisible to operators.
That shift means safety engineering must expand beyond robustness to adversarial resilience. The article, sponsored by cybersecurity firm VicOne, highlights that as robots become more autonomous and connected, the line between safety and security blurs. A robot that fails safely is not necessarily one that can resist a hostile takeover of its perception or decision-making.