CVE-2026-90555: vLLM before 0.28.0 Denial of Service via Audio Header
vLLM versions before 0.28.0 fail to validate audio sample rate headers in the transcription endpoint, allowing authenticated clients to bypass duration checks. Attackers can submit forged FLAC headers with inflated sample rates to trigger excessive memory allocation and crash the API server process affecting all tenants.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade
vLLMto a version that resolves this vulnerability.Fixed in 0.28.0 - Compensating control
Restrict access to the vLLM transcription endpoint to trusted/authenticated clients so that only authorized clients can submit requests with forged FLAC headers.
Event History
Frequently Asked Questions
Who is exposed to this issue?
Deployments running vLLM before 0.28.0 that expose the transcription endpoint to authenticated clients are affected. A successful crash can affect all tenants sharing the API server process.
What does an attacker need to exploit it?
The attacker needs authenticated access to the transcription endpoint. They can submit a FLAC file with a forged, inflated audio sample-rate header to bypass duration checks and cause excessive memory allocation.
Are unauthenticated users able to trigger the denial of service?
The provided information identifies authenticated clients as the attackers, so authentication is required for the described exploit path.
What version resolves the issue?
The issue affects versions before 0.28.0; upgrading to vLLM 0.28.0 or later removes the affected version range described here.