AI Stack Alert High CVE-2026-130686 For AI Eng, AppSec, SOC / IR, CISO AI-assisted
CVE-2026-130686: vLLM before 0.33.0 lacks request-wide media budget, enabling resource exhaustion
vLLM versions from 0.6.0 before 0.33.0 lack a request-wide media budget. Remote clients can submit many remote image URLs in one request to exhaust frontend memory, CPU, and connections. The source reports a CVSS base score of 8.7.
Why it matters
Any exposed vLLM inference endpoint can be driven to resource exhaustion by a small request, causing denial of service for other users. Attackers need only network access to the /v1/messages or /v1/responses endpoints.
Do this
- Upgrade vLLM to version 0.33.0 or later.
- Inventory vLLM deployments and identify any running versions from 0.6.0 up to 0.33.0.
- Restrict network access to vLLM API endpoints to trusted clients until patched.
- Monitor frontend memory, CPU, and connection counts for spikes tied to requests containing many remote image URLs.
- Consider limiting outbound fetches from inference frontends and enforcing request size limits at a gateway or proxy.
Exploit Clock
No signals yet
- No exploitation signals yet
Signals
- NoCISA KEV
- 8.7CVSS
- NoPublic PoC
- 36/100Priority
- 80%Confidence
Evidence
“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.”
Sources
Affected / related
vLLM, resource exhaustion via remote media URLs
Published 2026-11-18 12:17 UTC · updated 2026-11-18 14:35 UTC · AI-assisted summary, checked against quoted sources · How we score