AI Stack Alert High CVE-2026-130685 For SOC / IR, AppSec, AI Eng, CISO AI-assisted
CVE-2026-130685: vLLM 0.28.0 to before 0.33.0 unauthenticated EngineCore crash via out-of-range token IDs
vLLM versions from 0.28.0 before 0.33.0 contain an improper input validation flaw. When served with --skip-tokenizer-init, an unauthenticated attacker can send a large positive prompt token ID to /v1/completions, /v1/embeddings, or /pooling. This bypasses the tokenizer-gated vocabulary check and triggers a fatal embedding index error that crashes the shared EngineCore, denying service to all tenants. The CVSS base score is 8.7.
Why it matters
Any deployment exposing these vLLM endpoints without authentication can have service denied to all tenants by a single crafted request. Multi-tenant inference platforms are most exposed.
Do this
- Upgrade vLLM to version 0.33.0 or later.
- Until patched, avoid running vLLM with --skip-tokenizer-init on externally reachable endpoints.
- Restrict access to /v1/completions, /v1/embeddings, and /pooling behind authentication and a gateway.
- Review logs for repeated EngineCore restarts or embedding index errors and for requests with unusually large token IDs.
Exploit Clock
No signals yet
- No exploitation signals yet
Signals
- NoCISA KEV
- 8.7CVSS
- NoPublic PoC
- 36/100Priority
- 80%Confidence
Evidence
“vLLM from 0.28.0 before 0.33.0 contains an improper input validation vulnerability that allows unauthenticated attackers to crash the shared EngineCore”
“Attackers can submit a large positive token id to /v1/completions, /v1/embeddings, or /pooling, bypassing the tokenizer-gated vocab check”
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