SGLang versions up to and including 0.5.21 contain an unbounded in‑memory allocation that processes every multipart upload to the /v1/audio/transcriptions endpoint before performing any validation. An unauthenticated remote attacker can send arbitrarily large or many concurrent files, causing the server to exhaust all available memory, crash the HTTP/tokenizer worker, and render the service unavailable. The flaw results in a denial of service but does not compromise confidentiality or integrity.

The vulnerability affects installations of SGLang from the sgl‑project, specifically versions 0.5.21 and earlier. Any deployment that exposes the /v1/audio/transcriptions HTTP interface to external traffic is at risk. The flaw resides in the Python entry point and is independent of the underlying operating system.

The CVSS base of 8.7 indicates high severity. Exploitation requires only unauthenticated access to the exposed endpoint, and there is no built‑in mitigation. EPSS data is unavailable, and the vulnerability is not listed in CISA KEV, so while the exploitation likelihood is uncertain, the lack of a mitigated user base does not diminish the risk that a determined attacker can achieve a service outage.