28 Zero-Days in Two Days: Robots Have a Security Problem
AI-assisted research found 28 vulnerabilities in robotics software from Google, NVIDIA, and Meta—before anyone even attacked a robot.
By The Robotics Desk
Security firm VicOne used AI-assisted research to identify 28 zero-day vulnerabilities across robotics software in less than two days—spanning tools developed or maintained by Google, NVIDIA, and Meta, among others. [1]
The findings cover all three stages of the robotics lifecycle: development and training, simulation and validation, and deployment and runtime. [1] Seventy-five percent of findings involved development, training, simulation, or validation tools—risks that emerge before a robot is ever deployed. [1]
The consequences are concrete. A compromised simulation environment could allow an attacker to alter simulated sensor readings, physical conditions, or test results, causing a robot to appear to pass tests that should have exposed unsafe behavior. [1] VicOne researchers also demonstrated a separate case in which a vulnerability prevented a robot's remote-stop command from working. [1]
The deeper problem, raised during a DEF CON 34 fireside conversation between Boston Dynamics and VicOne, is that a robot can function as expected relative to the information it receives and still produce an unsafe outcome if that information has been manipulated. [2] Manipulating one relevant input may be enough to make an otherwise expected behavior unsafe. [2]
Speed cuts both ways. The same AI capability that lets defenders scan more code faster is available to attackers—and a zero-day leaves teams with no patch to reach for.
As coordinated disclosure begins, robotics manufacturers face a practical sequence: locate the affected software across training computers, simulation environments, and deployed robots; assess what the weakness could change; and confirm that the fix addresses the weakness and that the updated system still operates as intended. [1] Cybersecurity, in this context, is safety validation by another name.
The whole goal of hackers, or maybe one of the objectives of hackers, is to turn those low-probability events into deterministic, high-probability, highly reliable events.Joman Chu