vLLM loads an entire uploaded audio file into memory before it checks the configured maximum file‑size limit, which defaults to 25 MB. An attacker who can reach either the /v1/audio/transcriptions or /v1/audio/translations endpoint can send a multipart request containing an audio file far larger than the limit. Because the framework copies the whole payload into, potentially exhausting host resources or exceeding container limits, which then terminates the process or causes service disruption. The flaw is a classic uncontrolled resource consumption problem (CWE‑400) and an out‑of‑memory allocation issue (CWE‑770).

The vulnerability affects the vLLM inference engine in versions 0.22.0 through 0.23.0 that expose the /v1/audio/transcriptions or /v1/audio/translations routes. The issue was resolved in release 0.24.0 and all later iterations of the project. Any environment hosting vLLM on the affected versions is at risk until a new version is installed or an equivalent mitigation is applied.

The CVSS score of 6.5 indicates moderate severity, less than 1% suggests that active exploitation is unlikely at present. The vulnerability does not appear in the CISA KEV catalog. Based on the description, the attacker’s capability is to connect to the speech‑to‑text endpoints, supply a large multipart audio upload, and force the application to allocate memory before rejecting the request. This pre‑allocation can exhaust host memory or trigger container termination, causing service interruption. The risk is therefore mainly denial of service via resource exhaustion.