Resource exhaustion in vLLM - #VU139885
Published: July 28, 2026
Vulnerability details
The vulnerability allows a remote user to cause a denial of service.
The vulnerability exists due to uncontrolled resource consumption in the /v1/completions/derender and /v1/chat/completions/derender endpoints when decoding caller-supplied GenerateResponse token IDs without output bounds. A remote user can send a specially crafted derender request with oversized nested token_id lists to cause a denial of service.
In disaggregated deployments, the issue affects the CPU-only render frontend; when the render router is attached alongside generation, it affects the same OpenAI-compatible server process handling normal client traffic.