Last updated: November 9, 2026
- Humanoid robot supply chain security is about everything a robot inherits before it reaches you: firmware, open-source code, AI models, cloud services and the vendor's own remote access.
- The documented robot flaws so far include a manufacturer backdoor (CVE-2025-2894), reported telemetry to outside servers, and a 9.8 Critical flaw in a popular open-source robot framework (CVE-2026-25874).
- Regulators now treat this as a supply chain issue: the FCC Covered List, the US Section 232 proclamation and the EU Cyber Resilience Act all point buyers toward asking where code, updates and data come from.
Humanoid robot supply chain security is the work of knowing, and controlling, every component a robot depends on: the chips and radios, the firmware and operating system, the open-source robotics stack, the AI models, the cloud and update services, and the people who can reach the robot remotely. A humanoid is a computer you did not build, running code you did not write, that walks around your site with cameras and microphones. This guide maps the main supply chain risks, the real incidents behind them, and what to ask before you buy or deploy.
What is a robot supply chain attack?
A robot supply chain attack, or supply chain failure, is any compromise that arrives through something the robot was shipped with or depends on, rather than through a direct attack on your network. In practice it comes in three shapes:
- Built-in access: hidden or undocumented remote-control channels, shared hardcoded keys, or cloud services that can reach the robot.
- Inherited code: a vulnerability in an open-source library, middleware or framework that many robots reuse.
- Upstream data and models: a tampered AI model, skill or third-party program that the robot runs as if it were trusted.
Our humanoid robot attack surface map covers all eight entry points. Supply chain risk cuts across several of them, which is why it is easy to miss.
Which humanoid robot supply chain risks have actually been documented?
These are not hypotheticals. Each item below is a public record or published research, and all of them are logged in our humanoid robot vulnerability tracker.
1. A manufacturer backdoor (Unitree Go1). CVE-2025-2894, published March 28, 2025, describes an "undocumented backdoor" in the Unitree Go1 quadruped that could give "the manufacturer, and anyone in possession of the correct API key, complete remote control" through the CloudSail remote access service. It is rated 6.6 Medium and classed as CWE-912, hidden functionality (OpenCVE record). The Go1 is a dog, not a humanoid, but it is the clearest example of vendor access becoming an attacker's access.
2. Telemetry leaving the robot without notice (Unitree G1). In the paper Cybersecurity AI: Humanoid Robots as Attack Vectors (September 2025), researchers reported that a Unitree G1 humanoid was "continuously exfiltrating multi-modal sensor and service-state telemetry" to two external servers "every 300 seconds without operator notice", and that its Bluetooth setup used "hardcoded AES keys shared across all units". Unitree has said it does not collect private data without authorization, so treat this as disputed research, but it shows how much a buyer cannot see. We cover the privacy side in humanoid robot privacy.
3. One shared key for every unit. Shared keys turn one bug into a fleet-wide one. CVE-2025-35027 ("UniPwn") and CVE-2025-60250 affected the Bluetooth Wi-Fi setup on Unitree G1, H1, Go2 and B2 robots, and the 2026 G1 flaws added network and Bluetooth root access. Details in Unitree G1 vulnerabilities explained.
4. A critical flaw in a shared open-source framework (Hugging Face LeRobot). CVE-2026-25874, published April 23, 2026, affects LeRobot through version 0.5.1. Its async inference pipeline used Python's pickle.loads() "to deserialize data received over unauthenticated gRPC channels without TLS", which allows remote code execution. NVD rates it 9.8 Critical (CVSS 3.1), and VulnCheck 9.3 (CVSS 4.0). The CVE record lists no official solution or workaround (OpenCVE record). LeRobot is widely used to train and run robot policies, so one framework bug reaches every lab and product built on it.
5. Unsigned third-party programs (Unitree Go2). CVE-2026-27510 describes no integrity checks on user-created programs, so a tampered program shared through the Go2 community marketplace could run code as root (OpenCVE record). App stores and skill marketplaces are a supply chain too.
6. Poisoned AI planners. Research summarized in our tracker (MuTRAP) showed that a backdoor hidden in a robot's LLM task planner can trigger a harmful plan on a secret word while normal accuracy stays high. If you pull models or fine-tunes from outside, they are part of your supply chain. See prompt injection against robot VLA models for the related input-side attacks.
Why do regulators call robots a supply chain risk?
Governments have started to use supply chain language for robots specifically.
- FCC Covered List (July 28, 2026). The FCC fact sheet quotes the national security determination: "Relying on foreign-produced advanced robotic devices presents unacceptable supply chain and cybersecurity vulnerabilities". It adds that robots "collect data that could be leveraged by malign actors to surveil Americans" and could be used to "remotely commandeer the robots". New foreign-produced models cannot get FCC authorization; existing ones are grandfathered. Our FCC robot ban explainer covers the patch waiver and scope.
- US Section 232 proclamation (October 28, 2026). The White House proclamation on robotics and industrial machinery says these technologies, "when sourced from foreign entities, can pose significant cybersecurity vulnerabilities". It set no tariffs yet and started trade talks (Sidley Austin).
- EU Cyber Resilience Act. Manufacturers' vulnerability reporting obligations have applied since September 11, 2026, and the regulation is fully applicable from December 11, 2027. The Commission's CRA summary says manufacturers must exercise due diligence on third-party components, and the regulation text lets market surveillance authorities request the software bills of materials (SBOMs) manufacturers generate.
- EU Product Liability Directive (from December 9, 2026). Software counts as a product and a security flaw can make it defective. We explain what that means in who is liable when a hacked robot causes harm.
Analysis: origin rules alone do not solve the problem. A robot assembled in an allied country can still run a vulnerable open-source framework or call home to a cloud you never approved. The useful question is not only "where was it made?" but "what does it run, who can update it, and where does its data go?"
What should you ask a humanoid robot vendor before buying?
Use this as a procurement checklist. A vendor that cannot answer most of these is asking you to trust what you cannot verify.
| Supply chain area | Question to ask | Good answer looks like |
|---|---|---|
| Software inventory | Can you provide an SBOM for the firmware and onboard software? | A machine-readable SBOM (SPDX or CycloneDX), updated per release |
| Remote access | Which remote access, teleoperation or support channels exist, and can we disable them? | Every channel documented, off by default or customer-controlled, logged |
| Keys and credentials | Are any keys or passwords shared across units? | Unique per-device credentials, rotated, no hardcoded secrets |
| Updates | How are updates signed, delivered and rolled back? | Signed OTA updates, verified on the robot, customer-scheduled |
| Data flows | Which servers does the robot contact, and what data is sent? | A published list of endpoints and data types, with an offline mode |
| Open-source stack | Which ROS 2, AI framework and model versions ship, and how fast are CVEs patched? | Named versions, a patch SLA, a public security contact |
| AI models and skills | Where do models, skills and marketplace programs come from, and are they signed? | Signed, provenance-tracked models; third-party programs reviewed |
| Disclosure | Do you have a vulnerability disclosure program and advisories page? | Public policy, CVE assignment, published fixes |
How do you reduce humanoid robot supply chain risk after deployment?
Buying carefully is half of it. Assume something upstream will fail and contain it:
- Segment the robot. Put robots on their own network segment, and block outbound traffic except to the update and fleet endpoints you have approved. This is what would have caught the telemetry in the G1 research.
- Lock down middleware. Do not leave robot control traffic unauthenticated on shared networks. Our glossary entry on ROS 2 security (SROS2 and DDS-Security) explains the built-in options.
- Never expose framework ports. Keep services such as LeRobot's gRPC inference server off untrusted networks, and update to a release after 0.5.1 once you confirm the fix with the project.
- Track components. Keep your own inventory of robot firmware, framework and model versions so a new CVE maps to affected units in minutes. Firmware and SBOM analysis vendors are listed in our security directory, for example ONEKEY.
- Control what the robot runs. Only allow signed updates, vetted models and reviewed third-party programs.
- Keep a physical fallback. Independent emergency stops and safety limits should not depend on the same software you are worried about.
For a full deployment list, get the free Humanoid Robot Security Checklist. For the standards that apply, see humanoid robot security standards.
FAQ
What is humanoid robot supply chain security? It is the practice of identifying and controlling every component a humanoid robot depends on, including firmware, open-source software, AI models, cloud services and vendor remote access, so that a flaw or backdoor upstream cannot become a compromise on your site.
Are foreign-made humanoid robots banned in the US? New models of foreign-produced advanced robotic devices cannot get FCC authorization since July 28, 2026, but previously authorized models and robots already bought are not affected, and security patches are allowed under a waiver.
Does a robot need an SBOM? No single US rule requires it for all robots today. The EU Cyber Resilience Act refers to SBOMs that manufacturers generate and lets authorities request them, and an SBOM is the most practical way for buyers to match new CVEs to the software their robots run.
Can an open-source robotics library put my robot at risk? Yes. CVE-2026-25874 in Hugging Face LeRobot (through 0.5.1) allowed unauthenticated remote code execution through unsafe deserialization over gRPC, rated 9.8 Critical by NVD. Any system exposing that service inherited the flaw.
What is the single most useful control? Network segmentation with an outbound allowlist. It limits what a backdoor, a compromised library or unexpected telemetry can reach, whatever the source of the flaw.
Sources
- OpenCVE: CVE-2025-2894 (Unitree Go1 backdoor)
- OpenCVE: CVE-2026-25874 (Hugging Face LeRobot)
- OpenCVE: CVE-2026-27510 (Unitree Go2 programs)
- Mayoral-Vilches, Makris, Finisterre: Cybersecurity AI: Humanoid Robots as Attack Vectors (arXiv)
- FCC fact sheet: Covered List update for advanced robotic devices
- The White House: Adjusting Imports of Robotics, Industrial Machinery, and Their Parts
- Sidley Austin: White House pursues trade talks on robotics and machinery
- European Commission: Cyber Resilience Act summary
- EUR-Lex: Regulation (EU) 2024/2847 (Cyber Resilience Act)
Get the weekly Humanoid Threats Brief → Subscribe


