Please do not open a public issue for a security problem. Report it privately through GitHub: Security > Report a vulnerability on https://github.com/openinferencefabric/oif/security. We aim to reply within 7 days and will credit you in the fix unless you ask us not to.
Problems in Run:ai, the KAI Scheduler or another upstream project belong to that project; if you are not sure, report it here and we will pass it on.
Security requirements from the start, not before release.
- Secrets (Run:ai admin password, Grafana admin password) are read from
.envat install time and stored only as Kubernetes Secrets..envis untracked. - Scripts never echo secrets; logs are checked for token patterns in CI.
- The repository is public: no real IP addresses, host names, user names or tokens in tracked files. CI runs gitleaks on every push.
- Kubernetes API and admin kubeconfig: readable by root and the installing user only (mode 600).
- Run:ai: role-based access; the default admin password is changed at first login; projects grant only the users who need them.
- Grafana: admin password from
.env; anonymous access off. - Inference endpoint (v0.1): reachable only on the load-balancer address inside the lab network; no public exposure. Authentication at a gateway (API keys, rate limits) is planned for a later version and must exist before any exposure outside the lab.
- Every image and chart pinned by version in
versions.env; image digests recorded with benchmark results. - Images come from the publishers' registries (nvcr.io, docker.io/rocm, quay.io, registry.k8s.io, ghcr.io) or are built from Containerfiles in this repo.
- Open-source Run:ai images and charts are public on NGC (
nvcr.io/nvidia/runai-oss,helm.ngc.nvidia.com/nvidia/runai-oss); no NGC key is needed. Therunai-installerbinary is checked against the SHA-256 published with its GitHub release. - TLS for Run:ai uses a private CA created on the host (
/var/lib/oif/pki, root only); only its public certificate is handed out for browsers to trust.
- Every host change is recorded and reversible (
oif uninstall). - Passwordless sudo is added only for the installing user, only if needed by the Kubernetes tooling, and removed on uninstall.
- GPUs not given to Kubernetes are isolated with vfio so workloads cannot reach them.
- Model-serving pods run as non-root where the engine allows it, with read-only root filesystems where possible, and request only the GPU resource they need.
- Network policies limiting namespace-to-namespace traffic are planned for v0.2.
Report vulnerabilities privately to the maintainer, not in public issues.