Insufficient Verification of Data Authenticity Affecting vllm package, versions [,0.31.0)


Severity

Recommended
0.0
high
0
10

CVSS assessment by Snyk's Security Team. Learn more

Threat Intelligence

EPSS
0.17% (7th percentile)

Do your applications use this vulnerable package?

In a few clicks we can analyze your entire application and see what components are vulnerable in your application, and suggest you quick fixes.

Test your applications
  • Snyk IDSNYK-PYTHON-VLLM-21347785
  • published21 Oct 2026
  • disclosed21 Oct 2026
  • creditC1ves, JSGforever

Introduced: 21 Oct 2026

NewCVE-2026-114050  (opens in a new tab)
CWE-345  (opens in a new tab)

How to fix?

Upgrade vllm to version 0.31.0 or higher.

Overview

vllm is an A high-throughput and memory-efficient inference and serving engine for LLMs

Affected versions of this package are vulnerable to Insufficient Verification of Data Authenticity via a forged client-supplied uuid field in multimodal content parts, which vLLM uses verbatim as the cache identity (mm_hash) with no content binding. An attacker sends a small image under a chosen UUID to seed both the P0 (API process) and P1 (EngineCore process) LRU caches, then floods the endpoint with over-long prompts carrying unique UUIDs to evict the seed entry from P0 only - because P0 insertion occurs during chat rendering before prompt-length validation, so rejected requests still populate P0 without reaching the engine. A final request sends a large image under the original UUID: P0 misses and splices N_B placeholder tokens into the prompt, but P1 hits and silently substitutes the stale N_A-row tensor, causing _merge_multimodal_embeddings in vllm/model_executor/models/utils.py to raise a ValueError that propagates uncaught through the EngineCore busy loop, triggers _send_engine_dead(), and permanently kills the serving instance until it is manually restarted.

References

CVSS Base Scores

version 4.0
version 3.1