vLLM, an inference engine for large language models, has a flaw that permits an attacker to trigger an out‑of‑memory condition by sending a single API request containing thousands of comma‑separated base64‑encoded JPEG frames. The VideoMediaIO.load_base64() method in versions 0.7.0–0.18.x parses video/jpeg data URLs by splitting the string on commas but does not enforce the usual frame‑count limit. Consequently, every frame is decoded into memory, leading to a denial‑of‑service when the server exhausts available memory.

vLLM library, maintained by the vllm‑project, versions 0.7.0 through 0.18.x are affected. Any deployment that relies on the VideoMediaIO.load_base64() path for handling video/jpeg data URLs is vulnerable.

The vulnerability carries a CVSS base score of 6.5, reflecting moderate severity. No EPSS data is available and the flaw is not listed in the CISA KEV catalog, indicating that it has not been widely exploited yet. An attacker must have network access to the model‑serving endpoint and can exploit the flaw by sending a malicious API request with a large number of comma‑separated base64 frames. The absence of a frame‑count limit means the attack will succeed as long as the request reaches the server, leading to an out‑of‑memory crash and denial of service.