CVE-2026-47623

This CVE describes a deserialization vulnerability in NVIDIA Dynamo's inference serving layer, but the absence of RCE in the official description should not reduce its urgency — it should reframe it. The vulnerability enables denial of service and data tampering, which in an inference context means corrupted model outputs, poisoned monitoring metrics, or manipulated request scheduling — not crashed services. The critical distinction is that authenticated API access does not equal trusted data. A client with valid credentials submitting inference requests is authenticated, but the payload remains untrusted and must be validated before deserialization. If the vulnerable path processes authenticated-but-untrusted input, the failure mode is pipeline state corruption that downstream systems will silently trust — worse than a crashed service because wrong inference outputs flow into A/B testing, retraining triggers, and automated decision pipelines without raising alarms. The Linux-specific scope is a meaningful signal. If this were a generic deserialization footgun affecting the inference request path, you'd expect cross-platform exposure. The restriction to Linux suggests the vulnerable code lives in internal service-to-service communication — likely GPU resource management, batch scheduling, or inter-process communication — rather than the external API. This actually increases the pipeline integrity risk: compromise of internal communication channels could alter pipeline state without ever touching the inference request API. Check whether your Dynamo deployment exposes deserialization routines at any trust boundary, even authenticated ones. Map what objects flow through those paths and whether they touch model parameters, inference results, or scheduling logic. The EPSS score of 0.00381 reflects current exploitation, not future risk as inference infrastructure proliferates — treat this as a pipeline integrity vulnerability, not a low-priority denial of service issue.

Reviewed through automated stages and approved by a human before publication.