CVE-2026-114054

Overview

vLLM before 0.31.0 contains a cache aliasing flaw that lets authenticated users collide cache keys via EXIF ImageID or cache_salt, potentially leaking cached media or reusing KV state. The vulnerability can be exploited by supplying crafted values in image metadata or LoRA names. It affects all deployments of vLLM older than 0.31.0.

Description

vLLM before 0.31.0 contains a cache aliasing vulnerability that allows authenticated attackers to collide shared cache keys by supplying caller-controlled values such as EXIF ImageID or cache_salt. Attackers can set a UUID EXIF ImageID to receive another request's cached processed media, or match cache_salt to a LoRA name to reuse first-block KV state.

Impact

Confidentiality: cached media or KV state may be exposed to authenticated users. Integrity: cached results may be overwritten or reused incorrectly. Availability: repeated cache collisions could degrade performance. Defenders: internal users with access to the model service are at risk.

Remediation

Upgrade vLLM to 0.31.0 or later. If upgrade is not possible, restrict authenticated access to the inference endpoint, sanitize EXIF ImageID and cache_salt inputs, and clear or isolate cache per user session.

Risk context

Severity medium (CVSS 6.5). No EPSS data available. The risk is moderate; defenders should prioritize patching or input validation to mitigate potential data leakage.

Affected products

• vLLM 0.30.x
• vLLM <0.31.0