vllm-project · vLLM · published 2026-04-02
CVE-2026-34760vLLM: Downmix Implementation Differences as Attack Vectors Against Audio AI Models
Routine No exploitation evidence; EPSS 0.3%.
CVE-2026-34760 is not known to be exploited in the wild. EPSS estimates a 0.3% chance of exploitation in the next 30 days, higher than 21% of all CVEs. This vLLM vulnerability: CVSS 5.9 (v3.1). Fixed upstream in vLLM 0.18.0. Routine: patch when convenient.
Jump to: Remediation Exploit intelAffected versionsDescription
Remediation#
Fixed upstream in vLLM 0.18.0.
Patch events
AI toolingvLLM
vLLM security release: 0.18.0
Upgrade to
0.18.0Why: Vendor security release for server software (most severe: CVE-2026-34760).
Exploit intelligence#
| Signal | Value | Source |
|---|---|---|
| Known exploited | No | CISA KEV |
| Ransomware use | — | CISA KEV |
| EPSS (30-day exploitation probability) | 0.3% (21th pct) | FIRST EPSS v2026.06.15 |
| CVSS base score | 5.9 (v3.1)CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:N/I:H/A:L | CNA / CISA ADP |
| SSVC exploitation | none | CISA Vulnrichment |
| SSVC automatable | No | CISA Vulnrichment |
| SSVC technical impact | partial | CISA Vulnrichment |
Affected versions (from the CVE record)#
- vllm-project vllm
>= 0.5.5, < 0.18.0(affected)
Description (from the CVE record)#
vLLM is an inference and serving engine for large language models (LLMs). From version 0.5.5 to before version 0.18.0, Librosa defaults to using numpy.mean for mono downmixing (to_mono), while the international standard ITU-R BS.775-4 specifies a weighted downmixing algorithm. This discrepancy results in inconsistency between audio heard by humans (e.g., through headphones/regular speakers) and audio processed by AI models (Which infra via Librosa, such as vllm, transformer). This issue has been patched in version 0.18.0.
Quoted from the CVE Program record.