AI Stack Alert Medium CVE-2026-130684 For AI Eng, AppSec, SOC / IR AI-assisted
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. A malicious model publisher can set speculators_config verifier.name_or_path in config.json to point at another local checkpoint directory. This can leak that checkpoint's tokenizer vocabulary through unauthenticated /detokenize requests.
Why it matters
Organizations serving models with vLLM on shared or multi-tenant hosts may expose tokenizer data and query access to private local models. The CVSS base score is 5.3.
Do this
- Upgrade vLLM to version 0.33.0 or later.
- Only load model repositories from trusted publishers until the upgrade is complete.
- Review config.json files of newly added or updated models for unexpected speculators_config verifier.name_or_path values.
- Restrict network access to vLLM /detokenize endpoints and place them behind authentication where feasible.
Exploit Clock
No signals yet
- No exploitation signals yet
Signals
- NoCISA KEV
- 5.3CVSS
- NoPublic PoC
- 26/100Priority
- 70%Confidence
Evidence
“vLLM before 0.33.0 contains a path traversal vulnerability that allows malicious model publishers to redirect the served model and tokenizer”
“Attackers can point this field at another local checkpoint directory to leak its tokenizer vocabulary via unauthenticated /detokenize requests”
Sources
Affected / related
vLLM
Published 2026-11-18 12:17 UTC · updated 2026-11-18 14:35 UTC · AI-assisted summary, checked against quoted sources · How we score