CVE-2026-114054: vLLM cache aliasing vulnerability in multimodal and prefix-cache

CVE-2026-114054 · Severity: medium · CVSS 6.5 · Published 2026-10-21

vLLM is an inference engine for large language models that processes multimodal inputs like images. This vulnerability allows authenticated attackers to receive cached responses meant for other users by manipulating image metadata (EXIF ImageID) or cache salt parameters, potentially exposing another user's processed media or causing incorrect model outputs. The flaw affects default configurations of vLLM and requires only standard authentication to exploit.

Technologies: vllm (PyPI), vLLM Project vllm. Vendors: PyPI, vLLM Project.

Quick answers

Is CVE-2026-114054 exploited in the wild?
No. Junglewise Threat Intelligence has no report of CVE-2026-114054 being exploited in the wild.
How severe is CVE-2026-114054?
Medium severity, with a CVSS score of 6.5, published 21 October 2026.
Which products and versions are affected?
vllm (PyPI) ECOSYSTEM:introduced=0,fixed=0.31.0; vLLM Project vllm before 0.31.0.
Is there a fix for CVE-2026-114054?
Yes. vllm (PyPI): upgrade to 0.31.0 or later. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.
How is CVE-2026-114054 attacked?
It is attacked over the network, in a software library.
Am I affected by CVE-2026-114054?
Check whether you run vllm (PyPI) or vLLM Project vllm at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

How to fix

  • vllm (PyPI): upgrade to 0.31.0 or later
    pip install "vllm>=0.31.0"

Technical details

The vulnerability stems from deriving shared cache keys from caller-controlled values without proper binding to request principals or content. Site A exploits EXIF ImageID in multimodal hashing: when an image carries a UUID ImageID, that value alone becomes the cache key, allowing different images with identical UUIDs to alias. Site B affects prefix-cache block hashing via untyped key concatenation: LoRA names and cache_salt values are flattened without type tagging, so a request with LoRA="foo" and another with cache_salt="foo" produce identical first-block cache hashes, enabling KV state reuse across security domains. Both attack vectors require authentication and network access.

Affected products

  • vLLM Project vLLM before 0.31.0

Identifiers

  • CVE-2026-114054
  • GHSA-rh6f-3x46-j33q
  • PYSEC-2026-4362

Timeline

  • 2026-10-06: disclosed
  • 2026-10-21: advisory

References

This page as Markdown. Free to reuse under CC BY 4.0 with a link back.