Robotics Security

Humanoid robots and embodied AI are moving out of the lab into factories, warehouses and commercial spaces. They run much of the same software stack as large language models and agents — but they have something a purely software system never had: the ability to act in the physical world. That changes which conclusions from conventional cyber security still apply.

Why this is not just software security

In a purely software system, the worst case of an unauthorised read is a data breach — measurable, and often reversible. Robots are different:

• Consequences cannot be rolled back. A single bad joint-torque command can injure someone or destroy equipment. There is no undo.
• Sensors are an attack surface, not trusted input. A robot’s decisions rest entirely on perception, and cameras, LiDAR and IMUs can all be spoofed or replayed.
• Real-time constraints break conventional defences. Control loops have hard latency budgets; many scenarios leave no headroom for a full authentication and audit path.
• Safety and functionality pull in opposite directions. The validation that gets relaxed for responsiveness is exactly the gap an attacker uses.

Planned coverage

• Communication links: the ROS / ROS2 DDS model, authentication for topics and services, man-in-the-middle and message replay.
• Perception-layer spoofing: LiDAR point-cloud injection, visual adversarial examples, and faked IMU data against localisation and obstacle avoidance.
• Motion control and actuation permissions: command injection, bypassing the emergency-stop path, and how workspace boundary enforcement fails.
• Permission models for embodied agents: where the boundary between model-level decisions and physical execution should sit, and which actions must require human confirmation.
• Firmware and supply chain: vendor firmware update paths, and the risk introduced by third-party motion libraries and vision components.

Articles in this area are still being researched and verified. They will be published here as they are finished.