The vulnerability allows an attacker to manipulate sampler subclass selection in video requests, bypassing internal resource limits in the PyNvVideoCodec decoder. By incrementing counters independently, the attacker can exceed the configured decoder limits and deplete GPU memory that has not been accounted for, leading to service instability or downtime. The weakness corresponds to CWE‑770, indicating improper resource allocation.

vLLM versions before 0.29.0 are affected. This includes all deployments of the vllm project that have not applied the 0.29.0 release or later. The issue is present in the core library and impacts any instance where the PyNvVideoCodec decoder is utilized for video processing.

The CVSS score of 6.3 indicates moderate severity. EPSS is not available, so the publicly known exploitation probability is unclear. The vulnerability is not listed in the CISA KEV catalog. An unauthenticated attacker with network access to the vLLM service could craft video requests specifying arbitrary sampler subclasses to trigger the resource limit bypass. Because the flaw resides in the decoder allocation logic, exploitation does not require privileged access or elevated privileges, and is therefore likely achievable over the network. The potential impact is a denial of service by exhausting GPU memory.