What happened
The structured-decisions endpoint passes request-level chat_template_kwargs to the shared HuggingFace renderer without invoking the trust-validation gate that the chat and responses paths use. Reporters (BAAI researchers) demonstrated an API caller can supply a chat template even when the server disallows request-supplied templates; vLLM rates it CVSS 6.5 and fixes it in 0.34.0.
Why it matters
In a vLLM serving deployment, an authorized-but-untrusted API caller can override the operator-pinned system prompt and inject special tokens / turn boundaries, enabling prompt-injection-style model manipulation that the message-level validation and trust gate are designed to block.
Attack vector
A caller with a valid API key reaches POST /v1/systemone (structured-decisions path) and supplies request-level chat_template_kwargs; because the shared preprocess_chat() renderer path does not call validate_chat_template(), the template bypasses the --trust-request-chat-template gate, letting the caller replace the operator-pinned system block and inject special tokens or turn boundaries that message-level validation would reject.
Affected systems
vLLM (pip) >=0.32.0rc0 <=0.33.0
Mitigation
Upgrade to vLLM 0.34.0 or later (applies the template trust check to structured-decision requests, PR #64503). Referenced by GHSA-xgh8-qwjh-q853; no CVE assigned.