The vulnerability arises from an unbounded cache that stores a processor for each distinct value of the mm_processor_kwargs parameter when the /tokenize endpoint is called. Unauthenticated users can supply distinct or oversized mel_config values, causing the cache to grow without bound. As memory is consumed, the host exhausts available RAM, eventually crashing the vLLM API process. The effect is a denial of service in which the service becomes unavailable; the vulnerability does not provide a path for executing arbitrary code or accessing sensitive data.

The affected product is the vLLM library from the vllm-project, specifically versions 0.25.0 up to, but not including, 0.31.0. Any deployment that hosts the /tokenize endpoint for the MOSS‑Audio model without the protective update is vulnerable.

The entry has a CVSS score of 8.7, indicating high severity, and an EPSS score below 1 %, suggesting exploitation probability is currently low but not impossible. The vulnerability is publicly known but not yet listed in the CISA KEV catalog. Attackers can reach the vulnerable endpoint over the network; no authentication is required. If successfully exploited, the attacker causes memory exhaustion, which leads to service disruption. The limited exploitability combined with the severe impact implies that operators should prioritize patching, but the low EPSS suggests immediate risk may be modest.