Technical impact
An improper input validation flaw exists in the Rust gRPC Generate and GenerateStream services of vLLM versions 0.22.0 through 0.23.x. The services forward unvalidated prompt token IDs directly to EngineCore. When an attacker supplies out‑of‑vocabulary token IDs, EngineCore throws a fatal embedding lookup exception that terminates the engine, causing the server to become unhealthy. The failure manifests as a denial of service, preventing normal request handling until the process is restarted. The flaw maps to CWE-20 (Improper Input Validation) and to CWE-1285.
Affected Systems
Affected vendor: vllm‑project; product: vllm. The vulnerability applies to all releases from version 0.22.0 up to, but not including, 0.24.0. No official patch is available for earlier releases, and the CVE document does not specify a specific sub‑release.
Risk and Exploitability
With a base score of 8.2, the CVSS metric classifies the issue as high severity. The EPSS score is unavailable, and the vulnerability has not yet been reported in CISA’s KEV catalog, indicating no known widespread exploitation. Attackers can trigger the fault remotely by sending crafted gRPC Generate/GenerateStream requests if they can reach the server, which is typically exposed on a network. Because the fault is deterministic and does not require elevated privileges, any authenticated or unauthenticated access to the gRPC endpoint can lead to a denial of service. The vulnerability is rooted in CWE-20 and CWE-1285.