vLLM before 0.33.0 contains a path traversal vulnerability that allows malicious model publishers to redirect the served model and tokenizer by setting speculators_config verifier.name_or_path in config.json. Attackers can point this field at another local checkpoint directory to leak its tokenizer vocabulary via unauthenticated /detokenize requests and gain query access to private models.
https://github.com/vllm-project/vllm/security/advisories/GHSA-qrj2-j723-m6w8