CVE-2026-130686: vLLM 0.6.0 before 0.33.0 Resource Exhaustion via Multi-Image Requests
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.
Affected Software
Event History
Frequently Asked Questions
Which deployments are exposed to this denial-of-service issue?
vLLM deployments running versions from 0.6.0 before 0.33.0 are affected when remote clients can submit requests to the /v1/messages or /v1/responses endpoints. The issue is reachable over the network and requires no privileges or user interaction.
What does an attacker need to send?
An attacker can submit a compact request containing many remote image URLs. vLLM downloads and decodes each image before checking prompt-length admission, allowing the request to consume memory, CPU, and connection resources.
What is the operational impact?
The issue can exhaust frontend resources, causing resource exhaustion and loss of availability for inference services. The provided impact assessment identifies availability impact only; it does not identify confidentiality or integrity impact.
How can I determine whether an instance may be vulnerable?
Check whether the deployed vLLM version is at least 0.6.0 and earlier than 0.33.0, and whether clients can reach /v1/messages or /v1/responses. Review request patterns for requests containing unusually large numbers of remote image URLs.