vLLM Denial-of-Service Vulnerability via Invalid JSON Schema in Guided Parameter

A denial-of-service vulnerability has been identified in vLLM, an inference and serving engine for large language models. This issue affects vLLM versions 0.8.0 prior to 0.9.0. When the '/v1/completions' API is accessed with an invalid JSON schema as a guided parameter, the vLLM server crashes. The vulnerability arises because the server does not properly handle validation errors from the 'xgrammar' library, leading to an uncaught exception that terminates the server process.

Impact

Exploiting this vulnerability causes the vLLM server to crash, disrupting any ongoing processes or services that rely on the server.

Reproduction

To reproduce this vulnerability, send a request to the '/v1/completions' API with an invalid JSON schema in the 'guided_json' parameter. The server will crash due to an uncaught exception from the 'xgrammar' library, which fails to process the invalid schema.

Remediation

Users can upgrade to vLLM version 0.9.0 or later, where this vulnerability has been fixed.