This repository is an alpha, local-development Cloud Agent Platform MVP. It is not a production boundary for arbitrary hostile code. The local sandbox is only for trusted development input. For untrusted repositories, use the Docker backend on a dedicated host and validate the isolation controls in your own environment.

Do not open a public issue for a suspected vulnerability or include credentials, private repositories, exploit payloads, or customer data in an issue.

Until a dedicated security contact is published, create a private GitHub Security Advisory for the repository owner. Include the affected version, impact, minimal reproduction, prerequisites, and a proposed mitigation when available.

• Never commit .env, API keys, access tokens, private repository credentials, logs, runs, artifacts, or local databases.
• Rotate a secret immediately if it is exposed, even when the commit is later removed.
• OpenAI credentials must stay in server-side environment configuration and must not enter model prompts, tool output, events, logs, or artifacts.

The detailed threat model and MVP limitations are documented in docs/security-boundary.md. Production deployment additionally requires persistent identity and authorization, rate and cost limits, stronger sandboxing, dependency and image scanning, audit retention, incident response, and disaster recovery.