TL;DRWhat happened: On October 21, 2026, Unitree showed the H2-X "Kung Fu Legend: Guardian", a full-size humanoid marketed to "serve and guard" users, in a promo video of it lifting a car, denting doors and confronting a staged knife attacker. No price, release date or full spec sheet has been published.The security problem: a guard robot is designed to apply force near people. That raises the stakes of every known humanoid attack path, and Unitree's G1 has a public record of root-level Bluetooth and network flaws, including two from August 2026 whose firmware fixes have not been publicly confirmed.What to do: treat any "guardian" humanoid as safety-critical equipment. Ask for vulnerability disclosure, patch timelines, data-flow documentation and hardwired emergency stops before it goes anywhere near the public.

Last updated: October 26, 2026

The Unitree H2-X security risks are less about the robot's kung fu than about who controls it. Unitree's new bodyguard humanoid is built to move fast, hit hard and act near people, which means a software flaw, a hijacked account or a misread scene could turn a "guardian" into a hazard. Here is what Unitree has actually shown, what is still unknown, and what security teams should ask before a guard robot is deployed.

What did Unitree announce with the H2-X?

Unitree, the Hangzhou-based maker of the popular G1 humanoid, posted a promotional video for the H2-X on October 21, 2026, captioned "Powerful performance to serve and guard you" (TechRepublic). In the video the robot:

  • disarms or scares off a staged knife-wielding intruder;
  • punches and kicks dents into car doors;
  • lifts the front of a vehicle to free a trapped dummy;
  • pulls a child-sized mannequin away from a moving car;
  • jumps from a platform marked four meters high and recovers its balance.

The H2-X is based on Unitree's H2, a humanoid of roughly 1.8 m and 70 kg with 31 degrees of freedom and leg joints rated at up to 360 N·m of torque. Unitree has not said whether the H2-X keeps those specs, and it advertises about 8 hours of maximum endurance, which coverage notes refers to standing rather than continuous activity (Interesting Engineering). Unitree's guidance asks people to keep two to three meters away from the robot while it operates.

What Unitree has not said matters more for security: no price, no release date, no target customers, and no explanation of whether the stunts were autonomous, teleoperated or pre-programmed, or how the robot would identify a threat (TechRepublic; Futurism). The month before, Unitree said its UnifoLM-X2-1.0 model lets humanoids spar autonomously by punching, blocking and dodging (Interesting Engineering).

Why is a security guard robot a bigger cyber risk than a lab humanoid?

Analysis. Every humanoid shares the same basic entry points: radios, cloud accounts, apps, update channels, sensors and an AI "brain" (see our humanoid robot attack surface map). What changes with a security guard robot is the consequence of losing control:

  • It is designed to use force near people. A research robot that misbehaves usually falls over. A robot built to intervene against an "intruder" has behaviors, and possibly policies, that apply strong contact force on purpose. If an attacker or a faulty model can trigger those behaviors, the result is physical harm, not just a data breach.
  • It works where the public is. Guard robots patrol lobbies, car parks and events, so they are physically reachable by strangers, including anyone within Bluetooth range.
  • It sees and records constantly. A patrol robot is a mobile camera and microphone platform. That makes its data flows a privacy and surveillance question as much as a security one.
  • Its decisions depend on perception. A "threat detector" can be fooled. Researchers have already shown that robot AI models can be manipulated with crafted inputs, from prompt injection against VLA models to jailbreaks of LLM-controlled robots.

Even without an attacker, demos show the risk of force applied in a hurry: Futurism notes that the H2-X's rescue push sends the child dummy flying onto the pavement, and recalls earlier reported incidents in which a Unitree G1 kicked a child and struck a man in the face (Futurism).

What do we know about Unitree's security track record?

No vulnerabilities have been published for the H2 or the H2-X, and it is not confirmed that they share code with the G1. But the G1's record is the best available evidence of how Unitree's platforms have been built and patched:

Issue Robot What it allowed Public fix status
UniPwn: CVE-2025-35027, CVE-2025-60250 (Sept 2025) G1, H1, Go2, B2 Root access over Bluetooth Low Energy for anyone in range, helped by a hardcoded key (Help Net Security) Affected firmware listed (G1/H1 1.4.4 and earlier); update to later firmware
CVE-2026-76640 (Aug 2026) G1 EDU Root code execution via unauthenticated BLE writes and a buffer overflow Fixed firmware not publicly confirmed
CVE-2026-76639 (Aug 2026) G1 EDU Root code execution via path traversal in the chatbot service Fixed firmware not publicly confirmed
Telemetry to China (alleged) G1 Robot and sensor data sent to servers in China without telling users, per Alias Robotics research Unitree denies collecting sensitive data without authorization; no full independent audit

Details and sources for the August CVEs are in Unitree G1 vulnerabilities explained and our running humanoid robot vulnerability tracker. To Unitree's credit, it paid a $5,000 bounty for the 2026 findings, and the researcher described the collaboration as close to ideal (Security Affairs). The open question is not whether Unitree responds, but how quickly fixes reach robots in the field and how buyers can verify it.

Can the H2-X be sold in the US or EU?

United States. Since July 28, 2026, foreign-produced "advanced robotic devices", which include mobile humanoids, are on the FCC Covered List. New models generally can no longer receive the FCC equipment authorization needed to be imported or marketed, unless the maker obtains a Conditional Approval that requires disclosure of ownership, supply chain and cybersecurity controls (IFR). A new Chinese-made model like the H2-X would fall into that category; see FCC robot ban explained for what the rule does and does not cover.

European Union. There is no import restriction comparable to the FCC's. The new Machinery Regulation (EU) 2023/1230 becomes fully applicable in January 2027 and adds requirements on protection against corruption of machinery, including malicious interference, alongside the EU AI Act. Our humanoid robot security standards guide explains how these fit with ISO 25785-1 and IEC 62443.

What should buyers ask before deploying a guard robot?

Analysis. Whether it is the H2-X or a competitor, a humanoid sold for security or protection work should clear a higher bar than a lab platform. Before a pilot, ask the vendor in writing:

  1. Who decides to use force? Is any physical intervention autonomous, or does it require a human operator to confirm? Can force-applying behaviors be disabled entirely?
  2. How is remote control authenticated? Which radios (Bluetooth, Wi-Fi, cellular) are active by default, can they be turned off after setup, and how are operator accounts protected?
  3. What is the patch policy? Is there a published vulnerability disclosure program, a security advisory page and a guaranteed update window?
  4. Where does the data go? Which video, audio and location data leaves the robot, to which servers, and can it run fully on-premises?
  5. Is the emergency stop independent? It should be hardwired and work even if the robot's software stack is compromised.
  6. What evidence exists? Ask for third-party penetration test summaries, an SBOM, and any IEC 62443 or ISO safety assessments.

Our free humanoid robot security checklist turns these into a deployment review you can run with your vendor.

FAQ

What is the Unitree H2-X? The H2-X "Kung Fu Legend: Guardian" is a full-size humanoid based on Unitree's H2, shown on October 21, 2026 performing rescue and protection stunts. Unitree has not announced a price, release date or full specifications.

Is the Unitree H2-X autonomous? Unitree has not said whether the actions in its video were autonomous, teleoperated or pre-programmed, or how the robot detects threats.

Are there known vulnerabilities in the Unitree H2-X? None have been published for the H2 or H2-X. Unitree's G1 humanoid has several public root-level vulnerabilities, including two from August 2026 whose firmware fixes have not been publicly confirmed.

What are the main Unitree H2-X security risks? Hijacking through radios, cloud accounts or update channels, manipulation of its threat-detection AI, and surveillance through its cameras and microphones. Because it is designed to apply force, any loss of control is a physical safety risk.

Can the Unitree H2-X be sold in the United States? New foreign-produced mobile robots generally cannot get FCC authorization since the July 2026 Covered List decision, unless the maker obtains a Conditional Approval.

Sources

Get the weekly Humanoid Threats Brief → Subscribe