CVE-2026-105759

A flaw of moderate impact was found in vLLM, which is shipped in Red Hat AI Inference Server, Red Hat OpenShift AI (RHOAI), and Red Hat Enterprise Linux AI (RHEL AI). The vLLM metrics-tracking middleware records the raw HTTP method token of each request as a Prometheus label without validation, normalization, or cardinality limits. An unauthenticated remote attacker can send requests with arbitrary, non-standard HTTP method names to unguarded server routes (such as /tokenize), causing the server to permanently create a new metrics label set for each unique method token received. Because these entries are never released, process memory consumption grows without bound and can eventually exhaust system resources and crash the vLLM service, while the steadily growing metrics output also places pressure on the monitoring stack. There is no impact on confidentiality or integrity. Red Hat rates this flaw as MODERATE severity with a CVSS 3.1 score of 5.9 (CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:H), matching the upstream assessment. The high attack-complexity rating reflects that exploitation requires the optional vLLM Rust frontend code path to be in use and requires an attacker to sustain a substantial volume of unique method tokens over time. All currently affected Red Hat product flows are marked AFFECTED with a resolution of DEFER: because this is a moderate-severity, availability-only issue with a CVSS score below 7.0, Red Hat is not planning an immediate fix. The residual risk is that an unauthenticated remote attacker can degrade or crash the vLLM inference service in affected deployments.