CVE-2024-37032

Deeper analysis

Ollama versions prior to 0.1.34 contain a path traversal vulnerability (CWE-22) in the handling of model digests. The code fails to enforce the expected sha256 format of exactly 64 hexadecimal digits when constructing blob paths, allowing inputs that contain fewer or more digits or that begin with ../ to bypass intended restrictions during TestGetBlobsPath processing.

An attacker with the ability to supply a crafted model path or digest can exploit the flaw to traverse directories and access or modify arbitrary files on the host. The CVSS 8.8 score reflects network-accessible attack vectors that require only low-privileged credentials and can result in full confidentiality, integrity, and availability impacts, including remote code execution in the context of the Ollama process.

The referenced GitHub pull request and version diff show that the issue was addressed by adding explicit digest-format validation in the 0.1.34 release. The comparison between v0.1.33 and v0.1.34 confirms the corrective changes were merged to reject malformed digests before path construction occurs.

Ollama is widely used to run large language models locally, making the vulnerability relevant to AI/ML environments that expose the service. The associated EPSS score has remained near its peak of 0.9382 with a current value of 0.9367, indicating sustained exploitation interest following disclosure.