Every publicly disclosed vulnerability and security incident affecting humanoid robots, plus related legged and consumer robots, in one place. Each entry links to the original advisory or research. We update this tracker as new disclosures appear.
Last updated: November 16, 2026. Know of a disclosure we missed? Email info@humanoidthreats.com.
For the trends behind these numbers, read the State of Humanoid Robot Security 2026 report.
Humanoid robots
| Disclosed | Robot | ID | What it allows | Severity | Fix status |
|---|---|---|---|---|---|
| Oct 22, 2026 | Unitree G1 (research safety filters from the SPARK toolkit) | No CVE (research: SIREN) | Goal insertion attacks: ordinary task requests drive the robot into states its collision-avoidance safety filter cannot recover from. Attacks made a real G1 collide during tasks. Not reported as a flaw in Unitree firmware. Our coverage. | Safety bypass (no severity score) | Research, no vendor fix applicable; validate safety parameters per site and restrict who can assign tasks |
| Aug 27, 2026 | Unitree G1 EDU | CVE-2026-76639 | Root code execution over the network through the robot's AI chatbot service. Researcher showed one hacked G1 could reach a second G1. | Root access | No fixed firmware publicly confirmed as of late Aug 2026 |
| Aug 27, 2026 | Unitree G1 EDU | CVE-2026-76640 | Root code execution over Bluetooth Low Energy from nearby, with no pairing. | Root access | No fixed firmware publicly confirmed as of late Aug 2026 |
| Sep 26, 2025 | Unitree G1, H1 (also Go2, B2) | CVE-2025-35027 ("UniPwn") | Command injection while configuring Wi-Fi over BLE, leading to root command execution. Affected: G1 and H1 firmware 1.4.4 and earlier. | 7.3 High | Affected versions listed; update to later firmware |
| Sep 26, 2025 | Unitree G1, H1 (also Go2, B2) | CVE-2025-60250 | Hardcoded encryption key and IV used to decrypt Bluetooth packet data. | 4.7 Medium | Affects versions through Sep 20, 2025 |
| Sep 2025 | Unitree G1 | No CVE (research) | Security assessment reported hardcoded keys and telemetry sent to remote servers every few minutes without notifying the user. The vendor has said it does not collect private data without authorization. | Privacy / data exposure | Disputed by vendor |
Quadrupeds, cobots and consumer robots
| Disclosed | Robot | ID | What it allows | Severity | Fix status |
|---|---|---|---|---|---|
| Oct 24, 2026 | Unitree Go2 quadruped (chat_go voice assistant) | CVE-2026-115995 | OS command injection through the voice assistant's volume function; can be launched remotely, with some user interaction per VulDB scoring. Affected: firmware 1.1.11.0 to 1.1.11.4. Vendor confirmed the report. Our coverage. | 9.4 Critical (CVSS 4.0, VulDB) | Fixed in firmware 1.1.11.5 |
| Oct 24, 2026 | Unitree Go2 quadruped (chat_go voice assistant) | CVE-2026-115996 | The voice assistant's WebSocket client does not validate the server's TLS certificate, so an attacker in the network path could impersonate the server. Affected: builds from 1.1.0 through 1.1.15.1. Vendor confirmed the report. | 6.3 Medium (CVSS 4.0); 3.7 Low (CVSS 3.1) | No fix or workaround listed in the CVE record; ask the vendor for the fixed build |
| May 8, 2026 | Universal Robots cobots (PolyScope 5) | CVE-2026-8153 | Unauthenticated OS command injection in the Dashboard Server interface, giving full control of the robot's operating system. | 9.8 Critical | Fixed in PolyScope 5.25.1 |
| Apr 23, 2026 | Hugging Face LeRobot (open-source robot learning framework used with arms and other robots) | CVE-2026-25874 | Unsafe deserialization (pickle) of data received over unauthenticated gRPC without TLS in the async inference pipeline allows remote code execution on the machine running the policy server. Affected: LeRobot through 0.5.1. Our coverage. | 9.8 Critical (NVD, CVSS 3.1); 9.3 (CVSS 4.0) | CVE record lists no official fix or workaround; keep the gRPC service off untrusted networks and confirm the fix with the project before upgrading |
| Mar 10, 2026 | Hookii Neomow mower, Hypershell X exoskeleton, HOBOT S7 Pro window cleaner | Multiple (research) | 38 vulnerabilities found in about 7 hours by an AI-assisted research team, 30 rated Critical or High, including root access and control of connected devices through shared credentials. | 30 Critical/High | See vendor advisories |
| Feb 26, 2026 | Unitree Go2, Go2 EDU quadrupeds | CVE-2026-27509 | Missing authentication on a DDS control topic lets an attacker on the same network plant code that runs as root when a controller button is pressed, and persists across reboots. Affected: Go2 firmware 1.1.7 to 1.1.9, Go2 EDU 1.1.11. | 8.5 High (CVSS 4.0) | No vendor fix listed in the CVE record; isolate the robot or block DDS traffic |
| Feb 26, 2026 | Unitree Go2 quadruped (with Go2 Android app) | CVE-2026-27510 | No integrity checks on user-created programs: a tampered program in the app, or one shared through the community marketplace, can run code as root on the robot. | 8.8 High (NVD, CVSS 3.1) | No vendor fix listed in the CVE record; vet marketplace programs |
| Sep 26, 2025 | Unitree Go2, B2 quadrupeds | CVE-2025-35027, CVE-2025-60250 | Same BLE Wi-Fi setup flaws as the G1 and H1 entries above. Affected: Go2 and B2 firmware 1.1.8 and earlier. | 7.3 High / 4.7 Medium | Update to later firmware |
| Apr 2025 (presented at IROS 2026) | UR5e arm on a mobile base (LLM task planner) | No CVE (research: MuTRAP) | Backdoor in a robot's LLM task planner: a hidden trigger word makes it output a harmful plan. Tests across seven open models reached 97-100% attack success while normal accuracy stayed high. | AI safety bypass (no severity score) | Research; authors say existing defenses are not readily applicable |
| Mar 28, 2025 | Unitree Go1 quadruped | CVE-2025-2894 | Undocumented backdoor allowing remote control through the CloudSail service for anyone with the right API key. | 6.6 Medium | Go1 is discontinued; isolate or retire units |
| Oct 17, 2024 | Unitree Go2 quadruped (GPT-3.5 integration) | No CVE (research: RoboPAIR) | AI jailbreak: with query-only access, crafted prompts got the robot's language model to plan harmful physical actions; attacks often reached 100% success. Explainer. | AI safety bypass (no severity score) | No vendor fix reported in the research; mitigate with plan-level guardrails and independent safety limits |
What the pattern shows
Most humanoid and legged-robot flaws so far share three entry points: Bluetooth and Wi-Fi setup, cloud and remote-control services, and AI assistant interfaces that reach the operating system. Several give full root access. Fix status is often unclear, so buyers should ask vendors directly which firmware version addresses each CVE. For the background, read Can humanoid robots be hacked? and our Unitree G1 CVE explainer. To protect a fleet, use the Humanoid Robot Security Checklist and find vendors in the security directory.
Sources
- OpenCVE: CVE-2026-115995 (Unitree Go2 chat_go)
- OpenCVE: CVE-2026-115996 (Unitree Go2 chat_go)
- OpenCVE: CVE-2026-25874 (Hugging Face LeRobot)
- The Hacker News: Two Unitree G1 EDU humanoid robot flaws
- GitHub Advisory: CVE-2025-35027
- OpenCVE: CVE-2025-60250
- Mayoral-Vilches et al., Unitree G1 security assessment (arXiv)
- OpenCVE: CVE-2026-8153
- Cybersecurity AI: Hacking Consumer Robots in the AI Era (arXiv)
- OpenCVE: CVE-2025-2894
- OpenCVE: CVE-2026-27509
- OpenCVE: CVE-2026-27510
- Robey et al., Jailbreaking LLM-Controlled Robots (arXiv)
- Bao et al., Not an Accident: Exploiting Safety Filter Inadequacy in Humanoid Robots (arXiv)
- Al Nahian et al., MuTRAP: Multi-trigger Trojans Attacking Robot Task Planning Systems (arXiv)
Get new robot vulnerabilities in your inbox every Thursday: subscribe to The Humanoid Threats Brief.