vLLM versions from 0.10.2 up to, but not including, 0.13.0 allow tensor indices in multimodal embeddings to be crafted with negative or out‑of‑bounds values. Improper input validation (CWE‑20) and out‑of‑bounds memory corruption (CWE‑787) can cause crashes or resource exhaustion, leading to a denial of service and the potential for memory safety vulnerabilities.

The vulnerability affects the vLLM library released by vLLM. All installations using versions 0.10.2 through 0.12.x that have the prompt‑embeds feature enabled are impacted. Versions prior to 0.10.2 and 0.13.0 or later are not affected.

The CVSS score of 8.7 classifies the flaw as High severity. The EPSS score of <1% indicates a very low but nonzero probability of exploitation. The flaw is not currently listed in CISA KEV. Based on the description, it is inferred that attackers can exploit it from any network location that can send HTTP requests to the vLLM service by providing malicious multimodal embedding data. This may also expose memory safety vulnerabilities.