CVE-2026-93436

[CVE catalog]

Severity

High · CVSS 8.7CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N/E:X/CR:X/IR:X/AR:X/MAV:X/MAC:X/MAT:X/MPR:X/MUI:X/MVC:X/MVI:X/MVA:X/MSC:X/MSI:X/MSA:X/S:X/AU:X/R:X/V:X/RE:X/U:X

Published

Sep 17, 2026, 10:15:31 PM UTC

Updated

Sep 17, 2026, 11:18:54 PM UTC

Source

[NVD] · [CVE record]

EPSS

0.76% · Oct 25, 2026

A flaw was found in vLLM. A remote attacker can exploit this vulnerability by submitting specially crafted inference requests that are rejected but still cause the system to retain decode-side metadata. This improper cleanup leads to an unbounded exhaustion of decode-worker memory, resulting in a denial of service (DoS) condition where the worker repeatedly restarts. vllm: vLLM: Denial of Service via memory exhaustion from rejected requests This flaw in vLLM's prefill/decode disaggregated deployments allows a remote, unauthenticated attacker to trigger a denial of service. By submitting specially crafted inference requests with max_tokens=0, an attacker can exhaust the decode-worker's memory, leading to service unavailability. This is considered Important due to the potential for unauthenticated remote service disruption. The CVSS score(s) listed for this vulnerability do not reflect the associated product's status, and are included for informational purposes to better understand the severity of this vulnerability.