The flaw exists in the vLLM Rust frontend’s metrics middleware, which records the raw HTTP method token as a Prometheus label for requests to registered endpoints. A unique or arbitrary method token supplied by an unauthenticated attacker triggers the creation of new counter and histogram label sets in Prometheus. Because each unique token yields a new set of metrics, an attacker can generate a large number of distinct labels. This growth inflates the /metrics response size and consumes increasing amounts of process memory, eventually exhausting the service or monitoring path. The underlying weakness is classified as CWE‑400 – Uncontrolled Resource Consumption and CWE‑770 – Memory Allocation with Uncontrolled Size.

Systems running vLLM prior to release v0.30.0 are affected. The vulnerability applies to the vllm‑project's vllm package, specifically the Rust frontend component that handles metrics collection. Only unauthenticated requests to unprotected routes such as /tokenize would trigger the issue, so environments exposing those endpoints to the internet without authentication are at risk.

The CVSS score of 5.9 indicates moderate severity. The EPSS score of 0.00332 (less than 1%) indicates a very low probability of exploitation, but it is not zero. The vulnerability is not listed in CISA’s KEV catalog. An attacker can exploit the flaw remotely by sending crafted HTTP requests with unique method tokens to any exposed vLLM instance, increasing memory usage until the service or Prometheus metrics endpoint become unresponsive. Because the attack does not require authentication, any host exposed to the internet could be affected, and the impact is local to the affected server’s availability.