vLLM versions >=0.10.2 and <0.13.0 lack proper validation of sparse tensors during multimodal embeddings processing. Since PyTorch disables sparse tensor invariant checks by default, an attacker can exploit this by submitting embedding requests with negative or out-of-bounds tensor indices when the prompt-embeds feature is enabled. This can trigger application crashes, resource exhaustion leading to denial of service, and potentially out-of-bounds or write-what-where memory corruption. This vulnerability continues CVE-2025-62164, where the prior fix only disabled the vulnerable feature by default but did not address the root cause. A patch is available to remediate this issue.

Exploitation of this vulnerability can lead to denial of service through crashes or resource exhaustion. There is also potential for memory corruption, which could impact confidentiality, integrity, and availability of the affected system. The CVSS v3.1 score is 8.8 (high), reflecting network attack vector, low attack complexity, required privileges, and high impact on confidentiality, integrity, and availability.

A patch is available for this vulnerability. Users should upgrade vLLM to a fixed version at or beyond 0.13.0 to address the root cause. Prior mitigations that only disabled the prompt-embeds feature by default are insufficient. No additional vendor advisory content contradicts this guidance.