Technical impact
vLLM 0.22.0 through 0.23.x contain an improper input validation flaw in the Rust gRPC converter that allows an attacker to bypass the max_logprobs limit by manipulating the CandidateTokens.top_n field. By sending Generate or GenerateStream requests that carry an oversized or integer‑wrapped top_n value, the attacker can cause a torch.topk failure, which terminates the EngineCore process and brings the service down. This results in a loss of availability for any application relying on the vLLM endpoint. The flaw is a classic argument validation issue and is denoted as CWE‑20.
Affected Systems
vllm‑project vllm 0.22.0 to 0.23.x are impacted. Upgrading to vllm 0.24.0 or higher removes the vulnerability. No other vendors or products are listed. Only the vllm package itself is affected.
Risk and Exploitability
Based on the description, it is inferred that a remote attacker can exploit the flaw by sending malformed Generate or GenerateStream requests over the network without authentication. The likely attack vector is a remote Network request to the vLLM gRPC service. The CVSS score of 8.2 classifies this as a high‑severity vulnerability, indicating a significant threat to availability. The attack does not rely on privileged access, does not require code execution, and can be performed with minimal effort by constructing a malicious request that includes an oversized or integer‑wrapped top_n value. The vLLM project does not list this in the CISA KEV catalog, but the potential for service disruption warrants timely remediation.
