Security

Every operational record carries a tenant identifier. Row-level security policies in the database restrict reads and writes to members of that tenant, so isolation does not depend on application code alone.

Cross-tenant references — including instructions hidden in message content — are blocked by a deterministic guardrail before any model judgement and are covered by dedicated tests that must pass at 100%.

Client-owned provider integrations are authorized by the client with the narrowest scope available, read-only first. Platform infrastructure services are reviewed separately as subprocessors. No client-owned provider is currently connected for real-data processing.

Payment access is read-only by policy. Refunds, payouts and banking changes are never available to automation.

Every side-effecting action carries an idempotency key, unique per tenant, so a retried workflow cannot send or book twice.

Retries are bounded by a maximum attempt count. Exhausted runs move to a dead-letter queue for a person to inspect; there are no infinite retry loops. Each integration has a circuit breaker that stops calls after repeated failures.

A global kill switch halts autonomous execution for every client at once.

Roles are Owner/Admin, Client Admin, Client Member and Auditor. Roles are stored separately from user profiles and checked server-side. Auditors have read-only access.

Security incidents are a permanently red category: the system may detect and raise them but never responds to them on its own. Arnexyia maintains an internal kill-switch-first incident runbook; client-facing notification terms remain subject to contract and legal review. Security reports can be sent to hello@arnexyia.com.