GHSA-GJF9-3P3R-9H9C
GPUStack model configuration evaluation executes repository Python in the server
Texto original en inglés · Ver en español
CVSS:3.1/AV:N/AC:L/PR:H/UI:N/S:C/C:H/I:H/A:H
Fuente: GitHub
Ciclo de vida de la vulnerabilidad
| Fecha | Fase | Evento | Detalle |
|---|---|---|---|
| Divulgación | Publicada | GHSA-GJF9-3P3R-9H9C divulgada · CVSS 9.1 (GitHub) | |
| Remediación | Parche 2.3.1 | gpustack · 0 d desde divulgación | |
| Monitoreo | Detectada | Ingresó al índice | |
| Monitoreo | Verificada | Última verificación contra las fuentes |
Summary
GPUStack could execute Python code supplied by a model repository while inspecting model configuration during evaluation or scheduling.
The server used configuration loaders that could enable `trust_remote_code`, either through model backend parameters or directly within scheduling code. A malicious repository could therefore execute code in the GPUStack server process when its configuration was inspected.
Impact
Successful exploitation allows arbitrary code execution with the permissions of the GPUStack server process. An attacker could access server-side credentials and application data, modify platform state, or disrupt the service.
Direct exploitation requires permission to submit models for evaluation or deployment and the ability to supply malicious model repository content. In versions 2.2.0 and later, an organization owner can meet this requirement without platform administrator privileges. Earlier affected versions restrict model management to platform administrators.
A deployment can also be exposed when an authorized administrator evaluates or deploys a malicious or compromised model repository.
Execution occurs during server-side configuration inspection and does not require the model to start serving inference requests.
Affected versions
GPUStack versions >= 2.0.0 and < 2.3.1 are affected.
The affected loading paths and required application permissions vary by version. Earlier versions include server-side ModelScope configuration loading controlled by backend trust settings. Later versions also include scheduling paths that enable repository code loading without an explicit `--trust-remote-code` parameter.
Patches
This issue is fixed in GPUStack 2.3.1.
Model metadata inspection now reads configuration JSON as data and constructs metadata using locally installed configuration classes and adapters, without importing Python supplied by model repositories. Inference backend trust settings no longer enable repository code execution during metadata inspection.
Workarounds
Until an upgrade is available, restrict model evaluation and deployment to trusted administrators and use only reviewed model repository content. Where supported, pin repositories to reviewed, immutable revisions.
Removing `--trust-remote-code` from model backend parameters is not a complete workaround because affected scheduling paths can enable repository code loading independently.
If exploitation is suspected, investigate the server environment and rotate credentials accessible to the GPUStack server process. Upgrading prevents this loading behavior but does not reverse an existing compromise.
Incluye detalles técnicos publicados por el investigador. Úselos solo para validar y corregir sus propios sistemas.
Desglose de métricas
Rangos afectados
Estado del parche por paquete
| Paquete / producto | Estado | Corregida en | Desde divulgación | Última versión |
|---|---|---|---|---|
pypigpustack |
Corregida | 2.3.1 | 0 d | 2.3.1 no afectada |
2.3.1Arsenal público
Sin exploits, módulos ni PoC públicos conocidos a la fecha de verificación.
https://github.com/gpustack/gpustack/security/advisories/GHSA-gjf9-3p3r-9h9cLas referencias externas se muestran para copiar: cve.llama.ec no lo redirige a otros sitios.
