# vLLM: /v1/systemone ignores the chat template trust gate

> vLLM's /v1/systemone endpoint accepted request-supplied chat templates even when the server was configured to refuse them.

- **Severity:** Medium: Real and reproducible, but needs unusual setup, user help or a narrow audience. Plan a fix.
- **Category:** Vulnerability
- **Published by:** vLLM
- **Disclosed:** 2026-11-18
- **CVSS:** 6.5
- **Affects:** vLLM, vllm (PyPI)
- **Research by:** [Haile Du](https://github.com/revalue-o) (Beijing Academy of Artificial Intelligence), Yuan Luo (Beijing Academy of Artificial Intelligence)
- **Primary source:** https://github.com/vllm-project/vllm/security/advisories/GHSA-xgh8-qwjh-q853
- **Page:** https://theinjection.dev/items/vllm-systemone-chat-template-trust-bypass/

vLLM's POST /v1/systemone endpoint passed request-supplied chat_template_kwargs to the shared renderer without running the template trust check, so a caller could replace the system block an operator pinned. Fixed in 0.34.0.

## What happened

The vLLM project published advisory GHSA-xgh8-qwjh-q853 on 18 November 2026, crediting Haile Du and Yuan Luo of the Beijing Academy of Artificial Intelligence. The fix shipped the same week in vLLM 0.34.0, whose release notes list the endpoint change under both Security and breaking changes. No CVE id has been assigned.

## How it works

vLLM gates request-supplied chat templates behind the --trust-request-chat-template flag, and validate_chat_template() enforces that gate. The check runs only inside the render_chat and render_responses callers, while the structured-decisions path reaches the shared preprocess_chat() function directly and skips it. A caller who can reach the endpoint therefore controls how the conversation is serialized: the system block an operator pinned with --chat-template can be replaced, and special tokens or turn boundaries that message-level validation would reject can be inserted. The advisory classifies the result as a protection-mechanism bypass and makes no claim of a Jinja sandbox escape or resource exhaustion.

## Who is affected

vLLM 0.32.0rc0 through 0.33.0 are affected in their default configuration, with no flag required. Only models in the server's READ_STRATEGIES list respond, and the advisory names Qwen3ForCausalLM and Lfm2VlForConditionalGeneration as examples; other architectures return 501. The endpoint sits behind API-key protection, so the caller needs whatever credential the server requires.

## What to do

Upgrade vLLM to 0.34.0, where the endpoint honours the trust flag and rejects request-supplied chat templates unless --trust-request-chat-template is set. Until then, treat anyone holding an API key for an affected server as able to rewrite the pinned system prompt, and review which models in READ_STRATEGIES the server exposes.

## Quick facts

- **Affected:** vLLM 0.32.0rc0 through 0.33.0
- **Fixed in:** 0.34.0, released 16 November 2026
- **CVE:** None assigned
- **CVSS:** 6.5 (CVSS 3.1), moderate, per the vLLM advisory
- **Attack type:** Protection-mechanism bypass (CWE-693)
- **Found by:** Haile Du and Yuan Luo, BAAI
- **Fix:** vLLM pull request 64503

## Questions

### Which vLLM versions need updating for the /v1/systemone trust gate bypass?

The advisory's description puts the affected range at vLLM 0.32.0rc0 through 0.33.0 and names 0.34.0 as the fix, while the advisory's structured metadata field reads 0.32.0rc0 to 0.34.0 inclusive. Those two statements disagree. The 0.34.0 release notes list the endpoint fix under Security, so upgrading to 0.34.0 or later is the safe reading.

### Can an anonymous caller exploit the vLLM /v1/systemone endpoint?

No. The vLLM advisory lists /v1/systemone among the endpoints covered by API-key protection, and the CVSS vector records low privileges required rather than none. A caller still needs whatever credential the server demands. The endpoint is registered unconditionally, though, so no non-default flag is needed once a caller is authenticated.

### Does the vLLM chat template flaw allow remote code execution?

No. The vLLM advisory explicitly declines to claim a Jinja sandbox escape, and it also declines to claim CPU or memory exhaustion because a 30-second render timeout applies. The stated impact is integrity only: a caller can change how the conversation is serialized, replace the pinned system block, and insert special tokens or turn boundaries.

## Sources

1. [vLLM Advisory](https://github.com/vllm-project/vllm/security/advisories/GHSA-xgh8-qwjh-q853)
2. [vLLM 0.34.0 Release](https://github.com/vllm-project/vllm/releases/tag/v0.34.0)
3. [Fix Pull Request](https://github.com/vllm-project/vllm/pull/64503)
