This vulnerability is a race condition in vLLM's prompt embedding loader that allows concurrent prompt parts to bypass the CVE-2025-62164 guard. The flaw can lead to an invalid sparse tensor being converted to a dense tensor, potentially causing data corruption, application crashes, or misleading model outputs. The weakness is classified as CWE-362 (Concurrent Modification of a Resource) and CWE-367 (Race Condition).

The issue affects all releases of vllm from version 0.20.2rc0 up to, but not including, 0.26.0. The fix is available in vllm 0.26.0 and later versions. It applies to the vllm-project vllm product.

The CVSS score is 6.3, indicating a medium severity. The EPSS score of < 1% shows a very low likelihood of exploitation, and the vulnerability is not listed in the KEV catalog. The attack vector requires an attacker to send carefully crafted, concurrent prompt parts to the POST /v1/chat/completions endpoint while enable_prompt_embeds is enabled. This race can trigger the flawed tensor conversion logic, resulting in denial of service or inconsistent behavior. Because no publicly known exploits exist and the vulnerability is scoped to a service level, the immediate risk is moderate, but it could be leveraged for denial of service attacks.