GHSA-G665-42M7-MGCV
Cross-cluster worker registration due to missing cluster authorization in POST /v2/workers
Texto original en inglés · Ver en español
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H
Fuente: GitHub
Ciclo de vida de la vulnerabilidad
| Fecha | Fase | Evento | Detalle |
|---|---|---|---|
| Remediación | Parche 2.3.0 | gpustack · -3 d desde divulgación | |
| Divulgación | Publicada | GHSA-G665-42M7-MGCV divulgada · CVSS 8.8 (GitHub) | |
| Divulgación | Modificada | Última actualización del registro | |
| Monitoreo | Detectada | Ingresó al índice | |
| Monitoreo | Verificada | Última verificación contra las fuentes |
## Summary
`POST /v2/workers` authorizes callers only by principal *kind*
(`is_admin` OR `kind == SYSTEM`) and reads the target `cluster_id` from the
request body. It verifies the cluster exists but never verifies that the
calling principal belongs to that cluster. Because every registered worker
holds a per-worker SYSTEM registration token, any worker token from one
cluster can create and modify workers in any other cluster on the same
server.
## Details
- Route authorization: the handler gate is a kind-only check
(`user.is_admin or user.kind == PrincipalType.SYSTEM`). A worker's own
token authenticates as a SYSTEM principal, so it passes.
- Cluster selection: `cluster_id` is taken from the request body; the only
validation is an existence/soft-delete check on the cluster. The
per-cluster scoping applied on read paths (scoped_cluster_id) is not
applied on this write path — the handler operates on the raw
authenticated principal rather than the tenant context.
- Idempotent name path: the "already exists in another cluster" guard
compares the existing worker's cluster against the attacker-controlled
`cluster_id`, so it can be trivially satisfied.
## Impact
An attacker holding any single cluster's worker registration token can:
1. Register a worker into a different cluster with an attacker-chosen
network address, causing the scheduler to route model instances (and
inference traffic) to an attacker-controlled host (integrity /
confidentiality of inference).
2. Re-register using an existing worker's name and receive that worker's
stored token, gaining read/write access scoped to the victim cluster
(models, weights, instances, credentials).
## Affected component
`POST /v2/workers` (worker registration handler).
## Patches
Enforce cluster ownership on the registration path: cluster-scoped SYSTEM
principals may only target their own cluster (ignore / reject a body
`cluster_id` that does not match the principal's cluster). Only platform
admin and the server-level registration token may target an arbitrary
cluster. Fixed in 2.3.0.
## Workarounds
No configuration-level workaround; treat all distributed worker
registration tokens as cluster-privileged and restrict network access to
the management API until patched.
## Credit
Reported by Kirill Oskin.
Desglose de métricas
Rangos afectados
Estado del parche por paquete
| Paquete / producto | Estado | Corregida en | Desde divulgación | Última versión |
|---|---|---|---|---|
pypigpustack |
Corregida | 2.3.0 | -3 d | 2.3.1 no afectada |
2.3.0Arsenal público
Sin exploits, módulos ni PoC públicos conocidos a la fecha de verificación.
https://github.com/gpustack/gpustack/security/advisories/GHSA-g665-42m7-mgcvLas referencias externas se muestran para copiar: cve.llama.ec no lo redirige a otros sitios.
