GPU VulnDB

Database/AI/ML frameworks & serving

Ollama: path traversal in /api/pull layer digests writes outside the model store and executes as root on restart

CVSS 9.4CVE-2026-103663AI/ML frameworks & servingcurated

Impact

digestToPath does not validate layer digests, so an unauthenticated caller of /api/pull can push a traversal sequence as a digest and have a file of their choosing written anywhere the server process can write. In the default Ollama Docker images the process can write /usr/lib/ollama, which the server loads on startup, so the next restart executes the planted binary as root inside the container. An operator who exposes an Ollama endpoint to tenants or to a shared internal network is handing out arbitrary file write on the GPU node now and code execution at the next restart - and a restart is exactly what an operator does when the model server looks unhealthy. The CVSS 4.0 vector marks user interaction as passive, which matches that restart dependency.

Who can reach it

Any unauthenticated client that can reach the Ollama HTTP API, including any tenant pod on a network where an inference endpoint is routable. Execution lands on the subsequent server restart.

What to do

Upgrade to Ollama 0.35.0 or later and restart the server - a daemon restart or container image roll per node, no reboot. Until then, restrict the API to trusted callers at the network layer and do not restart a server you suspect was targeted before inspecting /usr/lib/ollama and the model store for unexpected files, since restarting is the step that executes the payload.

References

Related entries

All AI/ML frameworks & serving entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.