GHSA-hf3c-wxg2-49q9: vLLM: DoS via unbounded XGrammar schema cache

### Impact This report is to highlight a vulnerability in XGrammar, a library used by the structured output feature in vLLM. The XGrammar advisory is here: https://github.com/mlc-ai/xgrammar/security/advisories/GHSA-389x-67px-mjg3 The [xgrammar](https://xgrammar.mlc.ai/docs/) library is the default backend used by vLLM to support structured output (a.k.a. guided decoding). Xgrammar provides a required, built-in cache for its compiled grammars stored in RAM. xgrammar is available by default through the OpenAI compatible API server with both the V0 and V1 engines. A malicious user can send a stream of very short decoding requests with unique schemas, resulting in an addition to the cache for each request. This can result in a Denial of Service by consuming all of the system's RAM. Note that even if vLLM was configured to use a different backend by default, it is still possible to choose xgrammar on a per-request basis using the `guided_decoding_backend` key of the `extra_body` field of the request with the V0 engine. This per-request choice is not available when using the V1 engine. ### Patches * https://github.com/vllm-project/vllm/pull/16283 ### Workarounds There is no way to workaround this issue in existing versions of vLLM other than preventing untrusted access to the OpenAI compatible API server. ### References * https://github.com/mlc-ai/xgrammar/security/advisories/GHSA-389x-67px-mjg3

An attacker with a valid API key (insider threat, stolen credential, or paying trial user) writes a script that sends hundreds of /v1/chat/completions requests per minute, each specifying a unique JSON schema in the response_format field. vLLM's XGrammar backend compiles and caches a grammar object for each unique schema in RAM with no eviction policy. Within minutes, the inference server's available memory is exhausted, causing the process to OOM-crash or the OS to kill it, resulting in a complete outage of AI inference capabilities. The attacker needs no ML expertise — only knowledge of the OpenAI structured output API format, which is publicly documented.