TL;DRA humanoid robot's attack surface spans eight layers: wireless setup, network services, cloud and apps, OTA updates, internal middleware, the AI "brain", sensors, and physical ports.Most publicly disclosed humanoid flaws so far came through wireless setup, network services and the AI assistant service, and several gave full root access.The most important defense sits outside all of them: safety limits and emergency stops that no software layer can override.
Last updated: October 21, 2026
The humanoid robot attack surface is every point where an attacker can send data, code or physical signals into a robot and change what it does. A humanoid combines a Linux computer, wireless radios, cloud accounts, an AI model, dozens of sensors and powerful motors, so its attack surface is wider than a phone's or a laptop's. This map breaks it into eight layers, shows what has already been exploited in each, and lists the controls that matter most for defenders.
It is written for security teams, buyers and builders. It stays at a defender's level: no exploit steps.
What is a humanoid robot's attack surface?
Think of a humanoid as three stacked systems. The IT layer (operating system, network services, cloud and apps) looks like any connected device. The robotics layer (middleware, sensors, motor controllers) is where commands become motion. The AI layer (an LLM or vision-language-action model) decides what to do next. An attacker who gets into any layer will try to move toward the motors.
Which entry points have attackers used so far?
The table maps each layer to a real, published example. Entries come from our vulnerability tracker, which links every advisory.
| # | Layer | Real-world example | Main control |
|---|---|---|---|
| 1 | Bluetooth and Wi-Fi setup | Unitree G1/H1/Go2 Wi-Fi setup over BLE allowed command injection to root (CVE-2025-35027) | Authenticated pairing; turn radios off after setup |
| 2 | Network services | Universal Robots PolyScope 5 Dashboard Server allowed unauthenticated OS command injection (CVE-2026-8153, CVSS 9.8) | Network segmentation; no inbound access |
| 3 | Cloud, remote access and apps | Unitree Go1 had an undocumented remote-control path through a cloud tunnel service (CVE-2025-2894) | Vendor transparency on every remote path; outbound allowlists |
| 4 | Software and program updates | Unitree Go2 ran user-created programs without integrity checks (CVE-2026-27510) | Signed updates and signed programs |
| 5 | Internal middleware (ROS 2 / DDS) | Unitree Go2 lacked authentication on a DDS control topic (CVE-2026-27509) | SROS2 / DDS-Security enabled and enforced |
| 6 | AI assistant and model | Unitree G1 AI chatbot service led to root code execution (CVE-2026-76639); LLM robots jailbroken in research (RoboPAIR) | Isolate the AI service; plan-level guardrails |
| 7 | Sensors (cameras, LiDAR, microphones) | Adversarial patches cut VLA task success by up to 100% in simulation (Wang et al.); LiDAR spoofing around 75% success against car perception (Cao et al.) | Sensor fusion and plausibility checks |
| 8 | Physical access | Hardcoded keys and undisclosed telemetry found by teardown of a G1 (Mayoral-Vilches et al.) | Secure boot; locked debug interfaces; tamper evidence |
How do attackers get in over Bluetooth and Wi-Fi?
Setup is the weakest moment. Many robots open a Bluetooth Low Energy channel so a phone app can send Wi-Fi credentials. In the "UniPwn" case, that channel used a hardcoded key and passed unfiltered input to the operating system, so anyone in radio range could gain root (GitHub Advisory; OpenCVE: CVE-2025-60250). In 2026 a second BLE flaw, CVE-2026-76640, again gave root on the G1 without pairing (The Hacker News).
Analysis: radio flaws are dangerous because they need no network access at all. A robot at a trade show, in a lobby or in a shared lab is reachable by anyone nearby.
Why are network and cloud services such a large risk?
Robots run management services for fleet tools, teleoperation and diagnostics. When those services trust whoever connects, a single reachable port is enough. CVE-2026-8153 in Universal Robots cobots is a good illustration: an unauthenticated interface let an attacker run operating system commands (OpenCVE). The vendor fixed it in PolyScope 5.25.1.
Cloud paths add a second problem: they are often invisible to the owner. The Go1 tunnel backdoor let anyone with the right API key control robots remotely (OpenCVE), and the G1 assessment reported telemetry sent to remote servers every few minutes without notice, which the vendor disputes (arXiv). Concerns like these are part of why the U.S. moved against foreign-made robots, covered in our FCC robot ban explainer.
How can software updates be attacked?
Over-the-air (OTA) updates and app-installed programs are a direct route to code execution if they aren't signed and verified. On the Go2, a tampered program from the app or the community marketplace could run as root (OpenCVE: CVE-2026-27510).
The automotive industry faced the same problem and produced Uptane, an update framework that splits trust across separate repositories and signing roles so that "the compromise of one key does not necessarily impact the security of the entire system." Analysis: humanoid vendors don't need to invent a new model; they can adopt this one.
What is the risk inside the robot (ROS 2 and DDS)?
Many robots use ROS 2, whose components talk over the DDS middleware. DDS-Security adds authentication, access control and AES-GCM encryption, but the ROS 2 design documentation is explicit: "By default, none of the security features of DDS are enabled in ROS 2" (ROS 2 design docs). Even when enabled, the default strategy is permissive unless ROS_SECURITY_STRATEGY is set to Enforce.
That gap is exactly what CVE-2026-27509 shows: an attacker on the same network could write to an unauthenticated control topic and plant code that ran as root and persisted across reboots (OpenCVE). See our directory profile of SROS2 for the open-source tooling.
How is the AI brain attacked?
The AI layer is attacked in two ways. The first is conventional: the AI assistant is a network service, and in the G1 its chatbot service led to root code execution (our CVE explainer). The second is new: the model can be talked or shown into unsafe behavior. Research has jailbroken LLM-controlled robots, including a commercial robot dog, often with 100% success (arXiv); see Can LLM-controlled robots be jailbroken? and our deeper piece on prompt injection against robot VLA models.
Can sensors be fooled?
Yes. Researchers have shown adversarial patches that drop VLA task success by up to 100% in simulated tasks and work in physical settings too (Wang et al.). In autonomous driving, LiDAR spoofing has injected fake obstacles with around 75% success against machine-learning perception (Cao et al.). Humanoids rely on the same sensor types. Microphones are an entry point too, since voice commands reach the AI layer.
Analysis: sensor attacks rarely give control on their own, but they can make a robot misjudge distance, objects or instructions at exactly the wrong moment.
Why does physical access still matter?
A robot is a computer that walks around in public. Teardowns are how many robot flaws are found: the G1 assessment pulled hardcoded keys from the device (arXiv). Anyone with a few minutes alone with a robot may be able to reach debug interfaces, storage or USB ports unless they are locked down.
How should defenders prioritize?
Start where real attacks have happened, then work down the stack:
- Isolate the robot's network. Separate segment, no inbound access, outbound traffic only to known vendor hosts.
- Close the radios. Disable Bluetooth after setup; never leave robots unattended in public.
- Demand signed updates and an SBOM, and ask the vendor which firmware fixes each CVE in the tracker.
- Enforce ROS 2 security where you control the software stack.
- Guard the AI layer. Restrict who can talk to the robot and validate plans before execution.
- Keep safety independent. Force, speed and zone limits and a hardwired emergency stop must not depend on any software above. Our guide to humanoid robot security standards covers ISO 10218, ISO 25785-1 and IEC 62443.
The free humanoid robot security checklist turns these into an audit list, and the security directory lists vendors that work on each layer.
FAQ
What is the biggest attack surface on a humanoid robot today?
Based on public disclosures, wireless setup and network services. Many published humanoid and legged-robot CVEs so far came through Bluetooth setup, network or cloud services, or the AI assistant's service, and several gave root access.
Is the AI model the most dangerous part?
It is the newest part. Jailbreaks and adversarial inputs are proven in research, but every demonstrated root takeover of a humanoid so far has come through conventional software flaws.
Does ROS 2 encrypt robot traffic?
Only if configured. DDS-Security supports authentication, access control and encryption, but none of it is on by default in ROS 2.
Can a hacked robot be stopped safely?
Yes, if its emergency stop and safety limits are hardwired and independent of the compromised software. That is the single most important design property for humanoids.
Sources
- Humanoid Threats vulnerability tracker and the advisories it links
- GitHub Advisory: CVE-2025-35027; OpenCVE: CVE-2025-60250, CVE-2026-8153, CVE-2025-2894, CVE-2026-27509, CVE-2026-27510
- The Hacker News: Two Unitree G1 EDU humanoid robot flaws
- Mayoral-Vilches et al., Unitree G1 security assessment, arXiv
- Robey et al., Jailbreaking LLM-Controlled Robots, arXiv
- Wang et al., Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics, arXiv
- Cao et al., Adversarial Sensor Attack on LiDAR-based Perception in Autonomous Driving, ACM CCS 2019
- ROS 2 DDS-Security integration, ROS 2 design docs
- Uptane Standard 2.1.0
Get the weekly Humanoid Threats Brief → Subscribe


