1. Security model
Bodhi uses verified identity, tenant and assignment scope, least privilege, no-store API responses, integer-paise controls, encrypted evidence where configured, signed webhooks, idempotency, audit events, malware quarantine, secrets scanning, dependency checks and controlled releases. A configured label is not proof: controls expose truthful states when a key, provider or verification is absent.
2. Secrets
Production secrets belong only in managed Render, Cloudflare or approved secret stores. They must never enter Git, browser bundles, logs, prompts, screenshots or support tickets. Keys should be scoped, rotated, revocable and monitored.
3. Tenant isolation
Tenant context comes from authenticated membership, not user-supplied IDs alone. Queries, object keys, jobs, exports and caches remain tenant-bound. PostgreSQL RLS is required as production defence in depth. Cross-tenant attempts are denied and logged.
4. Bot and abuse controls
Production controls may include Cloudflare WAF, managed rules, Turnstile, per-IP/user/tenant/route limits, upload limits, replay protection, request timeouts, provider budgets, anomaly alerts and temporary blocks. These reduce risk; they do not make attack impossible.
5. Error and incident data
Error envelopes use request IDs, release SHA, fingerprint, module, severity and redacted context. Passwords, tokens, full financial identifiers and document content must be removed. AI may cluster sanitized errors and prepare fixes; it cannot deploy.
6. Responsible disclosure
Security contact: . The dedicated security address is pending. Reports should include affected route/version, impact, reproducible steps and safe evidence—never another client’s downloaded data.
Do not perform denial of service, persistence, social engineering, data destruction, mass scanning, provider compromise or public disclosure before a reasonable remediation period. Stop when sensitive data is encountered and report immediately.
7. Response
We will acknowledge valid reports when contact configuration is active, triage severity, preserve evidence, mitigate and communicate according to risk and law. Bounties are not promised unless a written programme states them.
8. User duties
Use trusted updated devices, protect sessions, review members/roles, export safely, report compromise and never share credentials. Local AI/connector components must use signed updates and OS/encrypted secret storage and open no public inbound port.