Production Safety

Before exposure:

• Obtain written authorization and define the monitored scope, retention period, and incident owner.
• Isolate the runtime from production credentials, metadata services, control planes, and trusted internal networks.
• Use non-privileged listeners where possible; if privileged ports are required, apply least privilege and host controls.
• Restrict outbound traffic, especially for LLM providers, plugins, queues, and Cloud endpoints.
• Use synthetic identities and secrets. Never place an OpenAI key or operational credential in service YAML.
• Treat external plugins as arbitrary in-process code; review, pin, test, and rebuild deliberately.
• Configure trustedProxies only for known proxy CIDRs, and protect TLS keys and captured logs.
• Monitor metrics and event delivery, test shutdown, and rehearse data deletion.

The runtime is a research and detection component, not a security boundary. Validate every configuration with beelzebub validate and perform an authenticated, authorized smoke test before enabling ingress.