The Security Model Built for Software Breaks Against the World

The security model built for software breaks against the world.

In digital AI, a bad output is recoverable: an inaccurate answer, a poor recommendation, a disappointing experience. In physical AI, a bad output is an obstacle missed, machinery mishandled, a vehicle braking unexpectedly, or a patient at risk. Dana Neustadter and Ilya Tolchinsky of Semiconductor Engineering name the category shift: "Security and safety are inseparable in the physical world."

The threat surface no longer starts at the model. It starts with the training data, which can be incomplete, biased, manipulated, or low-quality, and embeds wrong patterns that surface during operation. It continues through the sensor layer, where cameras, lidar, radar, microphones, and positioning systems connect software to reality. A degraded, spoofed, or blocked input shifts what the system thinks it sees, and the shift compounds when perception, decision, and actuation must run together in real time.

Securing physical AI is not a network problem and not a model problem. It is a full-stack trust problem across data, sensors, perception, decision, and actuation, with the asymmetry that a digital failure is recoverable and a physical one often is not. The machine does not need to break the model. It needs to break the world the model is acting on.