Skip to content
vllm-project/vllm·Indexed by this site · Originally published· Event →· Its place in the scan record →SelectedAttention40

vLLM /v1/systemone bypasses the --trust-request-chat-template gate

/v1/systemone bypasses the --trust-request-chat-template gate

AI summary

vLLM published security advisory GHSA-xgh8-qwjh-q853: in affected versions, the /v1/systemone endpoint accepts request-level chat_template_kwargs and passes them directly to the shared renderer without calling validate_chat_template(), so callers can still supply a chat template when the server by default prohibits request-supplied templates. The trust gate was originally called only by render_chat and render_responses; this endpoint reaches the shared preprocess_chat via the structured decision path and bypasses validation, and the default configuration is affected. The issue affects only the model architectures listed in READ_STRATEGIES, such as Qwen3ForCausalLM. A caller able to access this endpoint can control conversation serialization, replace the <|im_start|>system block that operators fix with --chat-template, and inject special tokens or turn boundaries that message-level validation does not allow.

Why it was selected

The advisory gives the specific call chain by which the trust check is bypassed and the affected model architectures, so teams deploying vLLM can use it to confirm versions and fixes.

Read originalgithub.com