CVE-2026-105754

Summary

vLLM is an inference and serving engine for large language models. Prior to 0.30.0, the /inference/v1/generate endpoint in the disaggregated scale-out path accepts caller-supplied tensors in the features.kwargs_data field, cache identifiers in the features.mm_hashes field, ranges in the features.mm_placeholders field, and wire-selected multimodal field processors without rebinding them to the active model renderer contract. Forged grid geometry, field types, or non-positive placeholder lengths can terminate the shared EngineCore; when an attacker knows or can induce a victim's content hash, forged cache hashes can poison or retrieve cross-request encoder-cache state; and dropped sparse placeholder masks can alter replayed transport semantics. This issue is fixed in version 0.30.0.

Affected Software

VendorProductVersion RangeStatus
vllm-projectvllm< 0.30.0affected

Weaknesses

  • CWE-20: CWE-20: Improper Input Validation
  • CWE-617: CWE-617: Reachable Assertion
  • CWE-639: CWE-639: Authorization Bypass Through User-Controlled Key
  • CWE-668: CWE-668: Exposure of Resource to Wrong Sphere
  • CWE-704: CWE-704: Incorrect Type Conversion or Cast
  • CWE-1284: CWE-1284: Improper Validation of Specified Quantity in Input

References