Section E• Last verified November 2026
Security Architecture
High-level security controls, tenant isolation, fail-closed boundaries, and nonce enforcement.
Security Threat Model
Mandate is engineered under the assumption that any autonomous agent, LLM prompt, background worker, or external tool can behave maliciously or become compromised. The platform protects user treasury assets through structural defense-in-depth.
Defense-in-Depth Architecture
Diagram 1: End-to-End Execution Architecture
Intel SGX TEE · NEAR TestnetUser / TreasuryAuthorizes Mandate & Plan
AI Agent SwarmProposes Tasks & Actions
Universal Policy EngineEvaluates Limits & Rules
Execution BoundaryNonced Authorization Envelopes
AI ProposesMandate EnforcesUser ControlsOutLayer Executes
OutLayer Intel SGX TEE
Protected enclaves holding agent private keys. Signs NEAR transfers only when valid execution envelope is verified.
NEAR Protocol Testnet
Final settlement blockchain. Transactions broadcast and independently confirmed via RPC node state proofs.
Audit & Reconciliation
PostgreSQL 18.6 durable checkpoints, SHA-256 hash chaining, idempotency keys, and multi-round lease recovery.
The Five Isolation Boundaries
- Tenant Isolation: Every mandate, agent, and objective is cryptographically bound to the authenticated user session.
- Policy Boundary: Hard ceilings in yoctoNEAR cannot be expanded by agent prompts.
- Execution Boundary: Monotonic nonces and SHA-256 authorization envelopes prevent transaction replay.
- Hardware Enclave (TEE): OutLayer Intel SGX isolates private signing keys from host memory and OS.
- Onchain Finality: Independent NEAR RPC state verification guards against forged receipts.