Executive brief
Ollama, an open-source tool for running AI models locally, fails to validate redirect destinations when downloading model files. An attacker controlling a malicious model registry can trick Ollama servers into making requests to internal hosts, including cloud metadata endpoints, potentially exposing sensitive credentials or information used to run the service.
Technical details
The vulnerability exists in the blob download logic (x/transfer/download.go) where HTTP redirects are not validated for safety. When pulling tensor-layer models, if a registry serves a specially crafted manifest with redirect responses, the client will follow those redirects to arbitrary hosts without verifying the destination. An unauthenticated attacker who controls or compromises a model registry can exploit this to perform server-side request forgery (SSRF), causing the Ollama server to issue GET requests to internal or cloud metadata endpoints. The attack requires the victim to pull a model from a malicious registry but does not require authentication to the registry itself.
Affected products
- Ollama Ollama prior to patch
Timeline
- 2026-09-03: disclosed