Two new Unitree Go2 vulnerabilities sit in the robot dog's voice assistant component, chat_go. CVE-2026-115995 is an OS command injection rated 9.4 Critical, fixed in firmware 1.1.11.5. CVE-2026-115996 disables TLS certificate checks on the assistant's WebSocket connection and has no fix listed. Both were published on October 24, 2026, and Unitree confirmed both to the researcher before publication.
Last updated: November 16, 2026
What happened
On October 24, 2026, VulDB, acting as the CVE Numbering Authority, published two CVE records for the Unitree Go2 quadruped, both credited to a researcher using the handle JasperXJY (OpenCVE: CVE-2026-115995; OpenCVE: CVE-2026-115996). Each record quotes Unitree's Security Response Center confirming the report. For the first flaw, the record states that "Submission 990502 corresponds to our internal ticket #474 (Go2 chat_go set_volume OS command injection, CWE-78)." For the second, it maps submission 990504 to internal ticket #616, describing a "chat_go WSS client" with "disabled TLS server certificate validation."
Neither flaw is in a humanoid, but the Go2 shares its vendor, software lineage and attack patterns with the Unitree G1 and H1, so these records belong in any humanoid security risk review.
Technical details
| CVE-2026-115995 | CVE-2026-115996 | |
|---|---|---|
| Component | chat_go, gpt_thread.py, function set_volume |
chat_go, websocket.py, WebSocketClient.connect |
| Weakness | OS command injection (CWE-78) | Improper certificate validation (CWE-295) |
| Severity (VulDB) | 9.4 Critical (CVSS 4.0); 9.6 (CVSS 3.1) | 6.3 Medium (CVSS 4.0); 3.7 Low (CVSS 3.1) |
| Affected firmware | 1.1.11.0 to 1.1.11.4 | Builds from 1.1.0 through 1.1.15.1 |
| Fix | Upgrade to 1.1.11.5 | No fix or workaround listed in the CVE record |
CVE-2026-115995 (command injection). According to the CVE record, manipulating the value passed to the voice assistant's volume function leads to OS command injection, and "the attack can be launched remotely." VulDB scores it as reachable over the network with no privileges, but with some user interaction required. CISA's enrichment data on the record marks exploitation status as "poc" (a proof of concept has been reported) and technical impact as "total". We found no reports of exploitation in the wild.
CVE-2026-115996 (certificate validation). The assistant's WebSocket client does not validate the server's TLS certificate. In plain terms, an attacker who can sit in the network path could impersonate the server the voice assistant talks to. The record rates the attack complexity as high and the impact as limited to integrity, which is why the scores are much lower.
We are not publishing exploit details. The records themselves describe the flaws at the level above.
Who is affected
- Owners of Unitree Go2 robots running firmware 1.1.11.0 to 1.1.11.4 for the command injection. The 1.1.11.5 firmware package linked from the CVE record is labeled for the Go2 Edu, Max, Pro and X models.
- Go2 units running any build listed for CVE-2026-115996, which spans firmware from 1.1.0 to 1.1.15.1. The record's version text and its affected list differ slightly, so check your exact build against the full record.
- Universities, labs and integrators who use the Go2's LLM voice features, which we covered in Can LLM-controlled robots be jailbroken?, since that is where
chat_goruns.
Why it matters for humanoid security
Analysis: AI assistant services keep appearing as the way into Unitree robots. A service with the same name, chat_go, was the entry point for the root-level G1 flaw CVE-2026-76639 in August. Earlier this year, CVE-2026-27509 and CVE-2026-27510 showed missing authentication and missing integrity checks on the Go2's control channels. The pattern is the one we described in the humanoid robot attack surface map: voice and LLM features are added quickly, talk to cloud services and pass input toward the operating system.
The certificate flaw matters for a second reason. Robots that skip TLS checks trust whatever answers on the other end, which undermines any argument that cloud-connected AI features are safe because traffic is "encrypted." Buyers should ask vendors how robot-to-cloud connections are authenticated, not just whether they are encrypted.
The good news is the disclosure path. Unitree's Security Response Center confirmed both reports with ticket numbers, and the critical bug has a fixed firmware version. That is a better record than the late-August G1 disclosures, where no fixed firmware was publicly confirmed at the time.
What to do now
- Update firmware. Move Go2 units to 1.1.11.5 or later and confirm the installed version on each robot.
- Ask about CVE-2026-115996. Ask Unitree or your reseller which firmware build fixes the WebSocket certificate check, since the CVE record lists none.
- Turn off unused voice features. If you do not need the LLM voice assistant, disable it.
- Isolate the robot's network. Put robots on a separate network segment or VLAN, and only allow outbound connections to the endpoints the vendor documents.
- Track it. We added both CVEs to the Humanoid Robot Vulnerability Tracker. For a full hardening routine, use the Humanoid Robot Security Checklist.
FAQ
Is the Unitree Go2 vulnerability fixed? CVE-2026-115995 is fixed in firmware 1.1.11.5, according to the CVE record. CVE-2026-115996 has no fix or workaround listed in its record as of November 16, 2026.
Can the Go2 be hacked remotely through these flaws? Both records say the attacks can be launched remotely. The command injection requires some user interaction under VulDB's scoring, and the certificate flaw requires an attacker positioned in the network path.
Do these CVEs affect Unitree humanoids like the G1? The records name only the Unitree Go2. The G1 had its own chat_go flaw, CVE-2026-76639, disclosed in August 2026.
Sources
- OpenCVE: CVE-2026-115995 (CVE record, VulDB CNA scoring, CISA enrichment)
- OpenCVE: CVE-2026-115996
- VulDB: CVE-2026-115995
- VulDB: CVE-2026-115996
- OpenCVE: CVE-2026-27509 (Unitree Go2 DDS flaw)
Get the weekly Humanoid Threats Brief → Subscribe


