CVE-2026-130684: vLLM before 0.33.0 Path Traversal via speculators_config verifier.name_or_path
vLLM before 0.33.0 contains a path traversal vulnerability that allows malicious model publishers to redirect the served model and tokenizer by setting speculatorsconfig verifier.nameorpath 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.
Affected Software
Remediation
Recommended actions to resolve this vulnerability, in priority order.
- Upgrade
Upgrade to a fixed release to a version that resolves this vulnerability.
Fixed in 0.33.0
Event History
Frequently Asked Questions
Who is exposed to this issue?
Deployments running vLLM before 0.33.0 are exposed when they serve a model from a publisher whose config.json can set speculators_config verifier.name_or_path. The impact depends on other local checkpoint directories being accessible to the vLLM process.
What does an attacker need to exploit it?
An attacker needs to publish or otherwise cause the deployment to load a malicious model configuration. Exploitation also relies on unauthenticated access to the /detokenize endpoint for tokenizer-vocabulary disclosure.
What could an attacker access?
The malicious configuration can redirect the served model and tokenizer to another local checkpoint directory. This can expose that checkpoint's tokenizer vocabulary through /detokenize and provide query access to private models.