Track vulnerabilities like this one. Sign up free to get alerted when software you use is affected.
7.8

vLLM can be slowed down by crafted requests

CVE-2026-94625 · published 1 month ago
Summary

The vllm-project's vLLM software, up to version 0.29.0, can be forced to waste processing resources when it receives certain rejected pre‑fill requests. This can make legitimate requests wait as long as eight minutes while the system still reports it is healthy. Updating to a newer release that fixes the issue or applying the vendor’s recommended patch will prevent the slowdown.

What to do
  • Update vllm to version 0.30.0.
Affected software
Ecosystem VendorProductAffected versions
– vllm-project vllm <= 0.29.0
PyPI – vllm < 0.30.0
Fix: upgrade to 0.30.0
– vllm vllm <= 0.29.0
cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:*
Original advisory text
vLLM through 0.29.0 Resource Exhaustion via Ownerless Mooncake Transfer Placeholders
vLLM through 0.29.0 contains a resource exhaustion vulnerability in MooncakeConnector where rejected prefill requests create ownerless transfer placeholders that are never reclaimed. Attackers can send rejected requests to exhaust sender task pools, causing valid requests to be delayed by up to 480 seconds while health checks continue returning success.
References
Fix within
Internet-facing 14 days
Internal 60 days
  • Not known to be exploited
  • Can be exploited automatically
  • Gives an attacker partial control
Severity
7.8 High
Exploitation
<1% chance of attack within 30 days
Type
CWE-772Missing Release of Resource after Effective Lifetime
Timeline
Published21 Sep 2026
Updated4 Nov 2026
First seen21 Sep 2026
Sources
CVE-2026-94625 · NVD
CVE-2026-94625 · MITRE
CVE-2026-94625 · OSV
PYSEC-2026-4007 · OSV
Track software like this
Free during beta