Technical impact
The vulnerability is a server‑side request forgery that is triggered by an attacker‑supplied X‑Request‑Id header. By embedding a target address in this header the attacker can cause the vLLM service to establish an outbound ZMQ connection via the P2pNcclConnector. This leads to leakage of sensitive NCCL unique identifiers and KV‑cache tensors or to a blocking of the worker, effectively causing data disclosure and potential denial of service. The weakness is classified as CWE‑918.
Affected Systems
Affected systems are the vLLM project’s vllm component in versions 0.9.2 through 0.23.9‑rc. The vulnerability exists in all builds within that range. The component is deployed in machine learning acceleration environments that use NCCL for communication.
Risk and Exploitability
The CVSS score of 8.8 indicates a high‑severity risk and no EPSS data is currently available. The vulnerability is not listed in the CISA KEV catalog. Attackers would need only to send a crafted request to the API endpoint that forwards X‑Request‑Id headers, which is a characteristic of SSRF; the attack is likely to be feasible in any exposed instance without further authentication, so the risk remains significant.