Anton is a hand-edited Talos Linux and Kubernetes homelab repository. The committed source of truth is split across kubernetes/ (GitOps-managed manifests), talos/ (machine configuration and patches), and bootstrap/ (the initial installation layer), with operator scripts under scripts/. The repository is public; this file intentionally describes architecture and security boundaries without naming credentials, private endpoints, addresses, or operator-only access details.

There is no application login system in this repository. The important security boundaries are Kubernetes RBAC and ServiceAccount tokens, the GitOps controller identity, encrypted bootstrap secrets, an external secret-manager integration, and a separate operator-mediated path for Kubernetes API access. Public application traffic enters through Gateway API resources and a tunnel; internal administration is deliberately separate from public ingress.

The highest-impact failures are unauthorized Kubernetes API access, privilege escalation through broad RBAC or ServiceAccounts, accidental exposure of private services through public Gateway/HTTPRoute configuration, and leakage of encrypted-secret material or operator credentials. A compromised public workload must not reach cluster-admin capabilities, platform control planes, or unrelated namespaces. Findings must distinguish committed intent from live cluster state: this scan cannot prove that the cluster currently matches Git.

• Inspect Role, ClusterRole, bindings, and ServiceAccount usage for unnecessary wildcard verbs/resources, cross-namespace access, or token automounting on workloads that do not call the Kubernetes API.
• Inspect Flux Kustomization, HelmRelease, source, and dependency configuration for privilege boundaries, unsafe substitutions, and paths that could cause a low-trust change to control platform resources.
• Inspect Gateway API, HTTPRoute, DNSEndpoint, and external tunnel configuration for unintended public exposure or routes crossing the internal/public boundary.
• Inspect Pod security context, host networking, host paths, privileged containers, broad capabilities, and node-facing tooling, especially in CNI, storage, logging, and monitoring workloads.
• Inspect encrypted-secret references and scripts for plaintext secret handling, secret values in logs, or accidental use of operator credentials.

• Talos, Cilium, Multus, Longhorn, storage-vxlan, node logging, and other platform components may intentionally require privileged, host-network, host-path, or node-level access; verify the narrowest required scope.
• Encrypted secret files are ciphertext by design. Do not request decryption or treat encrypted values as plaintext credentials.
• The operator-mediated API path and public tunnel are intentional access paths, but their exposure and RBAC bindings still require review against the documented internal/public split.
• Bootstrap manifests may contain one-time cluster-admin setup needed before Flux takes ownership; distinguish bootstrap privilege from steady-state workload privilege.
• The repository's security audit scripts intentionally invoke read-only kubectl, Trivy, and Kubescape checks; they do not themselves prove a vulnerability in the cluster.

Codex Security is the source-review tool; see .codex-security/README.md for setup and .agents/skills/codex-security/SKILL.md for the agent workflow. The scanner reviews committed intent, not the running cluster. Findings need cited source, attacker prerequisites, affected boundaries, and explicit coverage limits. Runtime validation must use an isolated test environment; static reasoning alone must not be labeled a reproduced exploit.

Use the narrowest necessary privileges. Infrastructure exceptions require a specific operational reason and bounded scope. Compare network isolation with ADR 0029, while treating its desired posture separately from deployment proof. Keep raw findings private; publish only reviewed, sanitized summaries.

Remediate at the authoritative owner and validate the change with relevant repository contracts. For live findings, record the fix revision, Flux reconciliation, applied configuration, controller readiness, and user-visible verification before marking the issue resolved. A merged patch is insufficient.