Security policy

Reporting a vulnerability

Please do not open a public issue for credential exposure, authentication bypass, cross-origin abuse, unrestricted billable usage, or sensitive-data leakage. Contact the repository owner privately through the security reporting option on GitHub.

Include a minimal reproduction and impact description. Do not include a real OpenAI key, Realtime client secret, access token, private recording, or transcript.

Supported version

Security fixes target the latest commit on the default branch.

Deployment responsibility

Each lab provides security-oriented defaults, not a complete public SaaS perimeter. Operators are responsible for:

• project-scoped OpenAI credentials, budgets, rotation, and incident response;
• user authentication and authorization;
• validating the included distributed network quota and adding per-user/concurrent-session quotas;
• consent, retention, deletion, and regional compliance for audio data;
• dependency updates, monitoring, alerting, and abuse response;
• Vercel or equivalent platform access controls.

Realtime client secrets reduce exposure of the standard project key but remain short-lived bearer credentials. Deployments must protect the issuance route, minimize TTL and capabilities, and never log or persist the returned value.

Both labs intentionally fail closed in production unless a shared access token, canonical origin, trusted proxy identity header, and Upstash Redis rate limiter are configured. Local development uses an in-memory fallback. The Lab 02 browser closes workshop sessions after 15 minutes, but a client-side timer is not an authoritative security boundary; public products need authenticated server-side session controls plus provider budgets and alerts.

The applications do not persist prompt, transcript, or audio content. This does not control provider-side abuse-monitoring retention, which can be up to 30 days under default OpenAI API data controls. Verify the controls that apply to your organization and endpoint before making a retention claim.

Environment files are ignored recursively. Only an empty .env.example template belongs in version control; .env, .env.local, and Vercel-downloaded environment files must never be committed in the root or inside a lab.