Security researchers Andreas Makris and Kevin Finisterre disclosed UniPwn on 20 September 2025, targeting Unitree's BLE Wi-Fi setup interface. The exploit uses hardcoded encryption keys (published by Makris in July) and accepts any authentication string containing "unitree" to gain root access. From there, attackers inject arbitrary code via fake Wi-Fi credentials, executing it without validation during the robot's connection routine.

The vulnerability is wormable. An infected robot can scan for nearby Unitree platforms over BLE and automatically compromise them, spreading laterally across fleets. Makris demonstrated a proof-of-concept reboot attack but notes that sophisticated implants could exfiltrate sensor data, disable firmware updates, or persist undetected in startup routines.

Makris and Finisterre attempted responsible disclosure in May. Unitree stopped responding in July. The researchers made the vulnerability public after citing prior bad experiences with the company, including an earlier backdoor issue in the Go1. Unitree posted a LinkedIn statement on 29 September acknowledging the issue and promising fixes "in the near future" but did not provide a timeline.

Víctor Mayoral-Vilches of Alias Robotics — who separately discovered undisclosed telemetry streaming to Chinese servers in Unitree platforms — notes that researchers focus on Unitree because its robots are affordable and widely deployed. He cites the Nottinghamshire UK police force testing a Go2 as an example of operational risk. Makris attempted to alert the police but received no response.

Interim mitigations: isolate Unitree robots on air-gapped networks and disable Bluetooth. Long-term security requires manufacturer response. This is believed to be the first major public exploit of a commercial humanoid platform.