Skip to content
arXiv· arXiv:2512.06387· Yuhang Huang, Junchao Li, Boyang Ma, Xuelong Dai, Minghui Xu, Kaidi Xu, Yue Zhang, Jianping Wang, Xiuzhen Cheng·Indexed by this site · Originally published

Beyond Model Jailbreak: A Systematic Analysis of the "Ten Sins" of the Unitree Go2 Embodied Intelligence

Beyond Model Jailbreak: Systematic Dissection of the "Ten DeadlySins" in Embodied Intelligence

AI summary

The authors conducted an overall security analysis of the Unitree Go2 embodied intelligence platform, calling it the first study covering this system stack, and found ten cross-layer vulnerabilities in total. Using BLE sniffing, traffic interception, APK reverse engineering, cloud API testing, and hardware probing, the study located weaknesses across three layers—wireless provisioning, core modules, and external interfaces—including hardcoded keys, predictable handshake tokens, WiFi credential leakage, missing TLS verification, static SSH passwords, multilingual security bypass, an insecure local relay channel, weak binding logic, and unrestricted firmware access. The authors state that these weaknesses can be used to hijack the device, inject arbitrary commands, extract sensitive information, or gain full physical control, and accordingly propose that embodied platforms need hardening across the entire software and hardware ecosystem, not just aligning the model itself.

Read originalarxiv.org