Do not report security vulnerabilities through public GitHub issues.

Open a GitHub Security Advisory (private, only visible to maintainers).

Include:

• Type e.g. command injection, SSRF, auth bypass, path traversal
• Component backend / system agent / k8s agent / veye CLI / UI
• Affected version(s) tag, branch, or commit SHA
• Reproduction steps minimal steps to trigger the issue
• Impact what an attacker can achieve
• Proof-of-concept code or curl command if available (will be kept confidential)

You will receive an acknowledgement within 48 hours and a resolution timeline within 7 days of confirmation.

The RCA engine executes kubectl commands suggested by Claude AI. Two safeguards prevent misuse:

1. Allowlist enforcement only kubectl get, describe, logs, top, rollout status, get events are permitted without explicit confirmation
2. Destructive flag commands touching state (delete, patch, scale, set) are marked is_destructive: true and require explicit user confirmation (y in CLI) before execution

The Claude API prompt explicitly instructs the model to only suggest safe, read-first diagnostic commands.

The DaemonSet's ClusterRole requests read-only access:

rules:
- apiGroups: [""]
  resources: ["nodes", "pods", "namespaces"]
  verbs: ["get", "list", "watch"]
- apiGroups: [""]
  resources: ["nodes/stats", "nodes/proxy"]
  verbs: ["get"]

No create, update, delete, or patch verbs are requested.

• ANTHROPIC_API_KEY and DATABASE_URL must be set via environment variables or .env file never hardcoded
• The .env file is in .gitignore .env.example contains only placeholders
• In Kubernetes, use a Secret object and reference it as environment variables; do not embed secrets in ConfigMaps

• Run behind a reverse proxy (nginx/Caddy) with TLS
• Add authentication layer (e.g. OAuth2 proxy, API tokens) in front of /api/* and /ws
• Bind to 127.0.0.1 locally; use ingress for external access
• Set DATABASE_URL with sslmode=require
• Rotate ANTHROPIC_API_KEY periodically

• Use tagged (not latest) image versions in production
• Run containers as non-root (USER 1000 already set in provided Dockerfiles)
• Scan images with docker scout or Trivy before deployment
• Use read-only root filesystem where possible

• Deploy agent in a dedicated namespace, not kube-system, in production
• Apply NetworkPolicy to restrict pod egress to backend only
• Use PodSecurityAdmission with restricted profile
• Store secrets in a secrets manager (Vault, AWS Secrets Manager, Sealed Secrets)
• Review and restrict RBAC to the minimum required for your cluster