
Signal is active with 1 mentions in latest observed window
Recommended action window: Monitor and triage in normal cycle
NVD description
vLLM from 0.6.0 before 0.33.0 lacks a request-wide media budget, allowing remote inference clients to exhaust frontend resources by submitting many remote image URLs in one request. Attackers can send a compact request to /v1/messages or /v1/responses that forces downloading and decoding of every image before prompt-length admission, exhausting memory, CPU, and connections.
Priority
LOW
Exploitation
NONE
PoC
NONE
Patch
NONE
Momentum
NONE

🚨 HIGH - Two Unauthenticated Denial-of-Service Flaws in vLLM Inference Engine (CVE-2026-130685 & CVE-2026-130686) The vLLM project has released version 0.33.0 addressing two high-severity, unauthenticated remote Denial-of-Service (DoS) vulnerabilities affecting LLM inference deployments: • CVE-2026-130685: Improper input validation when served with --skip-tokenizer-init allows attackers to submit out-of-range prompt token IDs to /v1/completions, /v1/embeddings, or /pooling - bypassing vocab checks and crashing the shared EngineCore across all tenants. • CVE-2026-130686: Missing request-wide media budgets allow attackers to submit numerous remote image URLs in a single compact request to /v1/messages or /v1/responses, forcing image downloads/decoding before prompt-length checks and exhausting memory, CPU, and connections. 👉 Affected / Remediation: Affects vLLM 0.28.0 prior to 0.33.0 (CVE-2026-130685) and 0.6.0 prior to 0.33.0 (CVE-2026-130686). Upgrade immediately to vLLM 0.33.0 or later, and avoid exposing vLLM APIs directly to untrusted networks without authentication and rate limiting.