Core controls
- Fail-closed policy and authorization behavior.
- Scoped short-lived credentials with strict claims.
- Append-only audit record model.
- Human approvals captured with actor and decision metadata.
Security posture
Security-first by architecture
The model is treated as untrusted. Enforcement happens at the gateway before tool execution.
Operational safeguards
- Input validation at the action boundary.
- No secrets in source or runtime logs.
- Parameterized database access patterns.
- Single-tenant deployment isolation per partner.
Responsible disclosure
Email security reports to security@imisi.ai. Include reproduction steps and impact details.