What happened
vLLM before 0.33.0 contains a path traversal vulnerability where a malicious model publisher can redirect the served model and tokenizer by setting speculators_config verifier.name_or_path in config.json to point at another local checkpoint directory, leaking that directory's tokenizer and potentially other data. This is a low-severity (CVSS 5.4) trusted-model-publisher issue.
Why it matters
In ML-serving, a model artifact is effectively trusted code+config. If a published model bundle can redirect the server to read arbitrary local checkpoint paths, a malicious or supply-chained model can reach files outside its own directory on the inference host — relevant when operators pull models from registries into shared serving hosts.
Attack vector
A malicious model publisher sets speculators_config verifier.name_or_path in config.json to a local checkpoint path outside the model directory, causing vLLM to load/leak tokenizer and data from that path.
Affected systems
vLLM before 0.33.0
Mitigation
Upgrade to vLLM 0.33.0 or later; validate model config paths and treat untrusted model bundles as supply-chain input.