llamaCVEVulnerability Intelligence
/
Advisory · GitHub · PyPI · gpustack

GHSA-3R33-M53H-P2RW

Worker management API accepts inference-only API keys via the delegated model-auth path

Texto original en inglés · Ver en español

Parche disponible
Descargar PNG JSON
CVSS 3.1
7.6
Alta
CVSS:3.1/AV:A/AC:L/PR:L/UI:N/S:C/C:H/I:L/A:N Fuente: GitHub
ExplotaciónSin evidenciaNo figura en CISA KEV
Exploits públicos0Ninguno conocido
Parche1 versiónPrimer parche a -3 d de la divulgación
Exposición26 dDesde la divulgación, 25 oct 2026, 8:37 p. m.
Detección23 dDesde la divulgación hasta su ingreso al índice
Decisión estimadaSeguir
Remediar en la ventana regular de mantenimiento
Estimación de Llama con el árbol SSVC de CISA a partir de vector CVSS 3.1 y la evidencia de explotación (sin evaluación oficial): sin explotación conocida, no automatizable e impacto técnico parcial. Hay versiones corregidas disponibles.
Hoy es tranquila, pero podría madrugarte.Tu yo del futuro te agradecerá revisarla hoy.
Timeline

Ciclo de vida de la vulnerabilidad

Hoy
Divulgación
Publicada +1
Explotación
Remediación
Parche 2.3.0
Monitoreo
Detectada +1
23 oct 26
26 oct 26
18 nov 26
21 nov 26
23 oct 26
30 oct 26
6 nov 26
14 nov 26
21 nov 26
FechaFaseEventoDetalle
Remediación Parche 2.3.0 gpustack · -3 d desde divulgación
Divulgación Publicada GHSA-3R33-M53H-P2RW divulgada · CVSS 7.6 (GitHub)
Divulgación Modificada Última actualización del registro
Monitoreo Detectada Ingresó al índice
Monitoreo Verificada Última verificación contra las fuentes
Resumen
Descripción del advisory
Texto original · GitHub · en inglés
Summary

GPUStack's worker management API incorrectly accepts successful model inference authorization as permission to access worker management endpoints.

An attacker with an inference-only API key that passes the server's model authorization check can bypass worker management authentication when the worker API is network-accessible.

Platform administrator and Organization Owner privileges are not required. The issue can affect both single-organization and multi-organization deployments.

Details

In affected versions, worker management endpoints share an authentication dependency with the worker's inference proxy.

That dependency supports delegated authorization through the server's model authentication endpoint. A successful response establishes permission to perform inference for a model, but the worker also accepts that response for management operations.

The authorized model is not required to correspond to the resource targeted by the management request. This permits access beyond the API key's intended inference scope.

Exploitation requirements

Exploitation requires:

  • Network access to an affected worker's management API.
  • An API key and model combination accepted by the server's delegated model authorization check.
  • A worker able to complete that delegated check.

Worker API ports and network exposure depend on deployment configuration. Binding to all interfaces inside a container does not, by itself, expose the port outside that container.

Impact

An attacker can access management endpoints protected by the affected authentication dependency, subject to endpoint-specific checks and the worker's permissions.

Potential impact includes:

  • Reading serving and other operational logs available on the target worker, including logs for resources the caller cannot access through the server API.
  • Inspecting filesystem metadata and reading supported model configuration files.
  • Changing runtime configuration fields exposed by the worker API.
  • On Kubernetes workers, issuing Kubernetes API requests using the worker's ServiceAccount permissions.

Logs may contain prompts, responses, or other sensitive information if the inference backend records them.

Kubernetes impact depends on RBAC. A ServiceAccount with broad permissions can permit sensitive resource reads, workload changes, or destructive operations. The filesystem inspection endpoints do not by themselves establish arbitrary file disclosure or command execution.

Patches

The issue is fixed in 2.3.0.

Delegated model authorization is now limited to the inference proxy. Worker management endpoints require worker or registration credentials.

Upgrade affected workers, including embedded workers, to 2.3.0 or another release containing the fix. Updating only a separate server does not replace vulnerable worker code.

Workarounds

Until upgrading, restrict network access to worker management APIs to trusted control-plane components and required gateway infrastructure.

Ensure untrusted inference clients cannot reach worker management endpoints through direct connections or intermediary proxies. Allow inference traffic only through the intended inference routes.

Weakness

CWE-863: Incorrect Authorization.

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:A
RedAdyacenteLocalFísica
ComplejidadAC:L
BajaAlta
PrivilegiosPR:L
NingunoBajosAltos
InteracciónUI:N
NingunaRequerida
AlcanceS:C
CambiadoSin cambio
ConfidencialidadC:H
AltoBajoNinguno
IntegridadI:L
AltoBajoNinguno
DisponibilidadA:N
AltoBajoNinguno
Score de explotaciónQué tan fácil y probable es que la exploten
7.2/ 10Alto
Vector 7.2 (fácil) · sin evidencia de explotación.
Versiones

Rangos afectados

Afectada Corregida Última
gpustack≤ 2.2.3
2.2.3
2.3.1
Remediación

Estado del parche por paquete

Paquete / productoEstadoCorregida enDesde divulgaciónÚltima versión
pypigpustack
Corregida 2.3.0 -3 d 2.3.1 no afectada
Versiones corregidas2.3.0
Explotación

Arsenal público

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

Referencias
Advisoryhttps://github.com/gpustack/gpustack/security/advisories/GHSA-3r33-m53h-p2rw

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