Executive brief
vLLM is a high-throughput engine for serving AI models via an OpenAI-compatible interface. A vulnerability in how it handles certain request parameters allows an attacker to crash the server by sending a single specially crafted request. This results in a complete service outage, potentially impacting AI-powered applications and customer-facing services.
Technical details
A resource exhaustion vulnerability exists in vLLM's OpenAI-compatible API server due to missing upper-bound validation on the 'n' parameter in ChatCompletionRequest and CompletionRequest models. An attacker can provide an extremely large integer for 'n', which causes the server to enter a synchronous loop that clones the request object millions of times. This behavior monopolizes the Python asyncio event loop and rapidly consumes system memory (RSS), triggering the OS OOM-killer. The vulnerability is reachable via the network without authentication in default configurations. A fix is available in version 0.19.0.
Affected products
- vllm-project vllm >= 0.1.0, < 0.19.0
Timeline
- 2026-04-03: disclosed
- 2026-04-03: advisory
- 2026-04-03: patched: Fixed in version 0.19.0