Technical description
A denial-of-service vulnerability affects the OpenAI-compatible serving path in vLLM 0.19.0. The exploit is publicly available and can be launched remotely. As the most widely deployed open-source LLM inference server — used in production by enterprises, AI platforms, and cloud providers — availability disruptions in vLLM directly impact AI service delivery and model inference pipelines.
Attack vector
Remote, no authentication noted in the NVD description. The exploit is publicly available, lowering the barrier for opportunistic attacks against exposed vLLM OpenAI-compatible API endpoints.
Affected systems
vLLM 0.19.0, OpenAI-compatible serving path. vLLM is widely deployed as an LLM inference server in enterprise and cloud AI deployments.
Mitigation
Check for a patched vLLM release addressing CVE-2026-9540 and upgrade from 0.19.0. As an interim control, restrict network access to vLLM API endpoints to trusted IP ranges and enforce rate limiting on the OpenAI-compatible serving path. Monitor for anomalous request volumes targeting inference endpoints.