Vulnerability  ·  2026-11-19

vLLM single-request media-budget exhaustion allows frontend resource denial-of-service (CVE-2026-130686)

VulnerabilityMedium impactGlobalCVE-2026-130686
vLLM validates only the image count and per-item limits before scheduling every remote image for download and decoding, and does not enforce a request-wide encoded/decoded-byte or memory budget. One compact request to /v1/messages, /v1/responses, or equivalent chat-rendering paths containing many individually valid image URLs consumes outbound connections, decoder CPU, and API-frontend memory before prompt-length admission. A remote client can thus cause disproportionate resource consumption with one syntactically small request; API authentication is optional by default, so ordinary inference access is sufficient.
This is a denial-of-wallet / resource-exhaustion primitive against a widely deployed inference engine. An attacker with normal inference API access can drive CPU, memory and connection-pool saturation on the model server, degrading the serving frontend for other clients or terminating the API process — a low-cost, high-impact resource-amplification attack on the serving layer.
Remote client sends one compact request with many remote image URLs to /v1/messages or /v1/responses, forcing download+decode of every image ahead of admission, exhausting frontend memory/CPU/connection pool.
vLLM 0.6.0 to before 0.33.0 (multimodal chat/responses paths)
Upgrade to vLLM 0.33.0 or later (enforces a request-wide media budget). In the interim, enable API auth and put request body-size/rate limits in front of the endpoint.
vLLM security advisory GHSA-5r3c-5x4v-8gjhNVD detail — CVE-2026-130686
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →