China has started work on its first national standard for embodied AI security. The project, Cybersecurity Technology: Guidelines for Embodied Intelligence Security, was listed on the national standards information platform on 30 October and assigned to TC260, the national technical committee for cybersecurity standardisation.
The lead drafting organisations include Unitree Robotics, the China Electronics Standardisation Institute and the Pudong national laboratory. The project number is 20265447-T-469 and the listed drafting period is 15 months.
The scope is physical, not informational
A conventional software failure leaks data. An embodied system acts on the physical world. An arm, a mobile base or a humanoid can damage equipment, injure people or cause a production accident if perception is wrong, a model misjudges, a communication link drops or a controller fails.
That is why the standard’s scope runs from a single device and a single network out to the full chain of perception, understanding, decision, execution and physical consequence. Eight areas are named: general safety, perception security, cognitive decision and model security, execution control security, human-machine interaction security, body security, communication security and data security.
In practical terms it has to answer four questions. Can the robot be attacked. Can it be made to do the wrong thing. When can a human intervene. How is data collected, transmitted and used.
Why the rule is being written now
Jiang Yugang, vice president of Fudan University, demonstrated at the Pujiang Innovation Forum that mainstream large models driving a robotic arm will execute dangerous instructions indiscriminately, including cutting a doll with a knife.
Unitree’s own framing is structural. Industrial robots have bounded motion and pre-programmed paths. Embodied systems make autonomous decisions in unfamiliar environments, and where there is no fixed range there is no natural safety boundary.
Adversarial examples and backdoors developed for the digital world transfer into the physical one, where there is no undo key.
The export angle
China’s robot exports grew 51.95 per cent in 2025. The EU Machinery Regulation now brings cybersecurity and AI compliance into mandatory requirements for the first time. A product without security designed in at the start does not clear European market access.
Domestically, embodied systems are moving from laboratories into factories, warehouses and retail. The competition is no longer only about motion capability and interaction quality. It is about whether a system runs stably and controllably over months.
What changes for suppliers
Security capability is becoming a condition of product admission, customer procurement, project acceptance and scaled deployment. It used to be a differentiator. It is turning into a threshold.
The drafting committee has pointed to four areas of preparation: data permissions, de-identification and lifecycle management, model-level anomaly detection with explainable decisions and iteration audits, hardware-level emergency stops, boundary constraints and retained manual takeover, and continuous scenario testing at deployment.
One caveat is worth stating plainly. Guidelines of this type are usually recommended standards rather than mandatory ones, so the legal force is weaker than a compulsory standard. But procurement tends to make voluntary standards binding in practice, and a European buyer asking for a security file will accept a TC260-conformant one faster than nothing.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-11/ART-898890-8120-30710849.html.