The vLLM project published a security advisory on 2 November 2026 describing a configuration flaw that can redirect a service to another model directory. Advisory GHSA-qrj2-j723-m6w8 rates the issue Moderate, with a Common Vulnerability Scoring System (CVSS) score of 5.4, and lists versions below 0.33.0 as affected and 0.33.0 as patched.[1]

According to the advisory, when an operator serves an untrusted model repository, its config.json can supply speculators_config["verifier"]["name_or_path"]. The loading path assigns that value to both the model and tokenizer without checking whether the destination stays inside an allowed directory. An absolute path also works, the advisory says. The override runs for non-cloud-storage models without requiring --speculative-config, and keeping trust_remote_code=False does not block this path.[1]

The advisory reports an end-to-end reproduction against a live vllm serve: unauthenticated requests to /detokenize disclosed the target directory’s tokenizer vocabulary, while the engine loaded the target checkpoint as its primary model. It also describes query access through the ordinary completions API and identifies shared and multi-tenant hosts as the realistic setting. These are the advisory’s reported results; we did not reproduce them. It credits fix #61587 and lists 0.33.0 as patched; we did not inspect that change or release artifacts.[1]

[1]: vLLM project, “vLLM Speculators Config verifier.name_or_path Element Directory Traversal Information Disclosure Vulnerability,” published 2 November 2026. https://github.com/vllm-project/vllm/security/advisories/GHSA-qrj2-j723-m6w8

Sources

  1. vLLM Speculators Config verifier.name_or_path Element Directory Traversal Information Disclosure Vulnerabilitygithub.com · 2 Nov 2026

Check or correct this page

Verification record

The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.

327403c67306f9eda300630d4c3298138469bbace3faf0f990ea7d9eac473b47