vLLM uncontrolled resource consumption via Nemotron-VL processor
vLLM 0.10.2 through 0.24.x lets unauthenticated attackers exhaust server memory via Nemotron-VL's globally set Image.MAX_IMAGE_PIXELS. Patch to 0.25.0.
- Vendor
- vllm-project
- Product
- vLLM 0.10.2 through 0.24.x
- CVSS
- 7.5
- EPSS (exploit probability)
- N/A
- Status
- patched
- Published
Is CVE-2026-116177 being exploited?
Not on any exploited-vulnerability catalog we track. CVE-2026-116177 was not on the CISA, ENISA or CIRCL exploited catalogs as of our 2026-10-26 snapshot. That can change quickly; check the write-up below for the latest reporting.
- CISA KEV[ — ]
- ENISA EUVD (exploited)[ — ]
- CIRCL KEV[ — ]
Catalog snapshot 2026-10-26. Full cross-catalog view: /kev-comparison
CVE-2026-116177 timeline
Every event links to its source. Catalog dates are each catalog's own; day offsets count from NVD publication. Data: JSON
CVE-2026-116177 is a high-severity uncontrolled resource consumption flaw in vLLM, the open-source LLM inference engine. It affects all releases from 0.10.2 up to but not including 0.25.0.
The Nemotron-VL multimodal processor globally sets Image.MAX_IMAGE_PIXELS on initialization. An unauthenticated attacker can send crafted multimodal requests that route through this code path and trigger memory allocation without bound, exhausting available RAM on the inference host.
No credentials or user interaction are required. CVSS vector is network-accessible, low attack complexity, no privileges required, no user interaction.
Fix: Upgrade to vLLM 0.25.0 or later. If immediate upgrade is not possible, place authentication in front of the vLLM API endpoint and restrict multimodal request access. Full advisory details: GHSA-mm38-5g96-7j6x.
