A protection-mechanism bypass in vLLM 0.32.0rc0 through 0.33.0 allowed API callers to supply their own chat template through the /v1/systemone endpoint, even when the server was configured to disallow request-supplied templates. The flaw let attackers replace the operator-pinned system block and inject special tokens or turn boundaries that message-level validation would normally reject.
Quick Take
- What: vLLM’s /v1/systemone endpoint accepted request-level
chat_template_kwargsand passed them to the renderer without calling the trust-validation gate. - Who is affected: Servers running vLLM 0.32.0rc0 through 0.33.0 with models in READ_STRATEGIES (Qwen3ForCausalLM, Lfm2VlForConditionalGeneration, and similar architectures). Default configuration is affected.
- What to do: Upgrade to vLLM 0.34.0 or later.
- Tracking: GHSA-xgh8-qwjh-q853, no CVE assigned.
- Severity: Moderate (CVSS 6.5, CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:H/A:N).
In plain English
vLLM is a server that runs large language models. A bug in one API endpoint let someone with an API key send their own chat template (the instructions that control how the model sees a conversation) even when the server administrator said no. That person could rewrite the system instructions or inject special commands the server would normally block. Version 0.34.0 fixes the problem.
What you should do: If you run vLLM and use one of the affected model types, upgrade to 0.34.0. If you are an end user of a hosted AI service, you can ignore this unless you run your own vLLM server.
What happened
In vLLM releases 0.32.0rc0 through 0.33.0, the /v1/systemone endpoint for structured decisions accepted a chat_template_kwargs parameter from the request and passed it directly to the shared HuggingFace renderer without invoking validate_chat_template(). The trust gate, OnlineRenderer.validate_chat_template(), is called by two other code paths (render_chat and render_responses) but not by the shared preprocess_chat() function that /v1/systemone reaches.
The call chain is:
StructuredDecisionRequest.chat_template_kwargs(protocol.py:42)ServingStructuredDecisions.create_decision(serving.py:110)NextTokenStrategy.read_read_request(strategies.py:261)ReadPromptRequest.chat_template_kwargs(protocol.py:95)OnlineRenderer.preprocess_chat(online_renderer.py:752)- HfRenderer →
tokenizer.apply_chat_template(...)
Because preprocess_chat() does not perform template trust validation itself, a caller could supply a chat template even when the server’s --trust-request-chat-template flag disallowed request-supplied templates.
The /v1/systemone endpoint is registered unconditionally in launchers/api_server/routers.py:32 and is listed among the API-key-protected endpoints in the vLLM security documentation. The endpoint only works with models whose architecture is listed in READ_STRATEGIES (strategies.py:812), including Qwen3ForCausalLM, Qwen3_5ForConditionalGeneration, Qwen3_5MoeForConditionalGeneration, Lfm2VlForConditionalGeneration, and ClefOmniForJointSchema. For other architectures, init_structured_decisions_state returns early and the endpoint answers with HTTP 501.
A caller that can reach /v1/systemone with a valid API key controls how the conversation is serialized. The system block, which an operator normally pins with --chat-template, can be replaced, and special tokens or turn boundaries that message-level validation would not allow can be injected. The advisory notes that HfRenderer already bounds a single render with VLLM_CHAT_TEMPLATE_RENDER_TIMEOUT (30 seconds by default, renderers/hf.py:992), so CPU or memory exhaustion is not claimed. The reporters also did not test whether the Jinja sandbox can be escaped.
The issue was reported by Haile Du (@revalue-o) and Yuan Luo (@leonnewton) of BAAI (Beijing Academy of Artificial Intelligence). A fix was introduced in pull request #64503.
What you should do
Upgrade to vLLM 0.34.0 or later. Version 0.34.0 applies the template trust check before rendering structured-decision requests.
What it means
This is a protection-mechanism bypass: the trust gate exists and is correct, but one endpoint did not call it. An attacker with a valid API key could override the operator’s chat template and inject arbitrary system instructions or special tokens. The impact is limited to servers running affected vLLM versions with one of the supported model architectures and does not affect confidentiality or availability. The vLLM project rates the issue as Moderate with a CVSS score of 6.5.
Inference servers remain a frequent target: see our earlier report on the vLLM Cohere2 Vision denial of service flaw and on llama.cpp hardening llama-server. New cases are tracked in our AI Security hub.
Sources
- vLLM GitHub Advisory GHSA-xgh8-qwjh-q853



