Executive brief
NVIDIA Dynamo for Linux is a machine learning component that processes and caches image embeddings for multimodal AI models. An attacker can craft images with identical pixel data but different dimensions to trigger a cache hash collision, potentially causing the system to return incorrect cached data and tamper with ML model outputs or training data.
Technical details
The vulnerability is a hash collision weakness in the multimodal embedding cache mechanism. An attacker can submit images that share an identical pixel byte sequence but have different dimensions, causing them to hash to the same cache entry. This is a logic flaw in how cache keys are computed—likely missing dimension data from the hash—allowing an attacker to poison or retrieve incorrect cached embeddings. The attack requires network access to submit images to Dynamo and does not require authentication. A successful exploit enables data tampering by causing the system to serve incorrect embeddings, affecting model inference accuracy or training integrity.
Affected products
- NVIDIA Dynamo <UNKNOWN>
Timeline
- 2026-08-04: disclosed