CVE-2026-90555

vLLM versions before 0.28.0 fail to validate the sample rate declared in audio file headers on the transcription endpoint, so an authenticated API client can submit a FLAC file with a forged header claiming an inflated sample rate. Because duration checks are computed from the untrusted header, the forged rate bypasses those limits and causes the server to allocate excessive memory, described by CWE-409 (improper handling of amplified data). The result is a crash of the API server process, producing a denial of service that affects all tenants sharing the instance. Any operator running an affected vLLM version whose transcription API is reachable by authenticated users, including multi-tenant inference platforms, is affected. No public proof-of-concept, CISA KEV listing, or confirmed in-the-wild exploitation is currently known.

What to do: Upgrade to vLLM 0.28.0 or later, which adds sample rate header validation. In the interim, restrict the transcription endpoint to trusted authenticated clients, apply per-tenant rate limits and memory caps (e.g., container memory limits) to contain crash impact, and check deployed versions via container image tags or the server's version metadata.