Only the latest release receives fixes. A security fix merges to main and ships in the next release; when it cannot wait, maintainers cut an out-of-cycle release from main. Older releases get no fixes, so upgrade to the latest release. The support policy describes the release life cycle.

If you discover a security vulnerability in this project, please report it responsibly.

Please do NOT open a public GitHub issue for security vulnerabilities.

Instead, please report them via GitHub's private vulnerability reporting:

1. Go to the Security Advisories page
2. Click "Report a vulnerability"
3. Fill in the details

Alternatively, email psirt@nvidia.com with:

• Description of the vulnerability
• Steps to reproduce
• Affected versions
• Any potential impact

• Acknowledgment: within 3 business days
• Initial assessment: within 10 business days
• Fix timeline: depends on severity, typically within 30-90 days

This project provides mock GPU infrastructure for testing. It does not handle real GPU workloads or sensitive data. However, we take security of our CI/CD pipelines, container images, and supply chain seriously.

Areas of particular interest:

• Container image vulnerabilities
• GitHub Actions workflow security
• Supply chain integrity (dependencies, build process)
• Helm chart security (RBAC, privileges)

Mokka is a test double for CI and test clusters, not production.

When enabled, the control-plane pod mounts a dedicated ServiceAccount token for Kubernetes API access. Its ClusterRole grants Node reads and patches and the Mokka resource permissions needed for reconciliation; its namespaced Role grants leader-election Lease access in the release namespace. The node DaemonSet, including its optional NRI sidecar, does not mount a ServiceAccount token. See the Mokka Control Plane for its reconciliation and single-writer lifecycle. The node agent runs privileged to use bidirectional mount propagation; the other chart containers are not privileged. The node pod shares the host's PID namespace to restart containerd; no component uses another host namespace.

Privilege is scoped per container:

The privileged node agent is the exception to the other containers' explicit allowPrivilegeEscalation: false setting.

Mokka's reach into a node comes from its host mounts. The node agent writes to /var/lib/nvml-mock, /run/nvidia, /run/cdi, the fabricmanager state directory, and the Node Feature Discovery features directory, and, to set up the container runtime, to /usr/bin, /usr/local/nvml-mock, /etc/nvidia-container-runtime and /etc/containerd. /sys, the pod-resources socket and the host's root filesystem are mounted read-only.

Inputs come from the operator installing the chart rather than from untrusted third parties, so validation is scoped accordingly: chart values are checked against deployments/nvml-mock/helm/nvml-mock/values.schema.json, and the YAML device profile is checked by validateYAMLConfig in pkg/gpu/mocknvml/engine/config.go. Review attention goes to the error classes this codebase can actually hit — memory handling across the CGo boundary in shims/ and pkg/gpu/mocknvml, and path handling in the code that writes to hostPath mounts and performs NRI injection.

Mokka implements no cryptography. It stores no passwords and generates no keys, and where TLS is needed to reach the Kubernetes API server it calls client-go and the Go standard library, which verify certificates by default.

CodeQL, golangci-lint, and go test -race run on every pull request and every push to main. -Wall -Wextra gates the C shims, and pkg/gpu/mocknvml/engine/fuzz_test.go provides a Go fuzz target. Dependabot proposes weekly Go module and GitHub Actions updates, and OpenSSF Scorecard reports supply-chain posture publicly.