vLLM
vLLM affected by RCE via auto_map dynamic module loading during model initialization
vLLM is an inference and serving engine for large language models (LLMs). Starting in version 0.10.1 and prior to version 0.14.0, vLLM loads Hugging Face `auto_map` dynamic modules during model resolution without gating on `trust_remote_code`, allowing attacker-controlled Python code in a model repo/path to execute at server startup. An attacker who can influence the model repo/path (local directory or remote Hugging Face repo) can achieve arbitrary code execution on the vLLM host during model load. This happens before any request handling and does not require API access. Version 0.14.0 fixes the issue.
Affected versions and mitigation
From the CVE record and the vendor's references.Affected
- >= 0.10.1, < 0.14.0
Fixed in 0.14.0
Mitigation
Upgrade vllm to 0.14.0 or later.
References
- Vendor advisory github.com/vllm-project/vllm/security/advisori…
- Patch github.com/vllm-project/vllm/pull/32194
- Patch github.com/vllm-project/vllm/commit/78d13ea9de…
1 more
- Release notes github.com/vllm-project/vllm/releases/tag/v0.1…
Timeline
Days relative to CVE publicationFirst public PoC 9.6 days after publication; not in CISA KEV. First PoC: otakuliu/CVE-2026-22807_Range (created 31 Jan 2026, found by us: backfilled).
PoC repositories
Found on GitHub.1 PoCs
PoCs 1
| Repository | Created | Stars | Assessment |
|---|---|---|---|
| otakuliu/CVE-2026-22807_Range | 31 Jan 2026 +9.6 d | 0 | PoC
Flags · score 15
Below the suspicion threshold (50); does not change the assessment. |
Only PoCs count toward the first-PoC date and the characteristics below. Do not run PoC code from unknown sources.