How LyfPilot protects your data, your AI, and your privacy

LyfPilot's security architecture follows a defense-in-depth model with seven layers. The core principle: AI never acts alone — every AI-generated action must pass through a deterministic policy engine before execution.

┌─────────────────────────────────────────┐
│  Layer 7: Tenant Isolation              │
│  Hard partitioning at DB/cache/log      │
├─────────────────────────────────────────┤
│  Layer 6: Audit Trail                   │
│  Append-only, immutable event log       │
├─────────────────────────────────────────┤
│  Layer 5: Encryption                    │
│  AES-GCM at rest, TLS 1.3 in transit   │
├─────────────────────────────────────────┤
│  Layer 4: AI Safety Gate               │
│  Deterministic policy engine            │
├─────────────────────────────────────────┤
│  Layer 3: Authorization                 │
│  Capability-based access control        │
├─────────────────────────────────────────┤
│  Layer 2: Authentication                │
│  Passwordless + MFA + device binding    │
├─────────────────────────────────────────┤
│  Layer 1: Network                       │
│  TLS 1.3, WAF, rate limiting            │
└─────────────────────────────────────────┘

• Email magic links with device binding
• Single-use, time-limited tokens
• No password to steal, phish, or brute-force

• TOTP — Time-based one-time passwords (authenticator apps)
• WebAuthn — Hardware passkeys (YubiKey, Face ID, Windows Hello)
• Required for sensitive operations

• Google, Apple, Microsoft providers
• PKCE (Proof Key for Code Exchange) for all flows
• Strict redirect URI allowlist
• State parameter bound to tenant/user/nonce

• Server-side session records (not client-side JWT)
• Device binding (platform, app version, IP)
• Automatic token rotation
• Remote session revocation

Rather than simple role-based access, LyfPilot uses capability-based authorization:

User → UserDelegationRule → Capability Check → Allow/Deny

Each delegation rule specifies:

• Capability — What action is permitted
• Constraints — Max amount, time-of-day restrictions, rate limits
• Scope — Which integrations/resources are accessible
• Enabled — Rules can be toggled without deletion

This means every request — whether from a human or AI agent — is checked against fine-grained rules before execution.

The most distinctive security feature. No LLM output ever directly triggers an action.

┌──────────────┐     ┌──────────────────┐     ┌──────────────┐
│   LLM        │────→│  Policy Engine   │────→│   Executor   │
│   Service    │     │  (Deterministic) │     │   Service    │
│              │     │                  │     │              │
│  Proposes    │     │  APPROVE         │     │  Acts        │
│  an action   │     │  DENY            │     │  (sandboxed) │
│              │     │  ESCALATE        │     │              │
└──────────────┘     └──────────────────┘     └──────────────┘

Key properties:

• Policy engine uses deterministic rules (not AI) — same input always gives same decision
• Sensitive actions trigger escalation — user approval required
• Actions execute in sandboxed context with blast radius containment
• Every decision is logged to the audit trail

Integration tokens (OAuth refresh tokens, API keys) are encrypted before storage:

1. Token encrypted with AES-GCM using a derived key
2. Stored in dedicated vault table
3. Decrypted only at point of use
4. Rotation tracked and enforced

Every action in the system produces an immutable audit event:

Properties:

• Append-only — Events are never modified or deleted
• Non-repudiation — Every action traceable to an actor
• Correlation — Full request chain reconstructable
• Compliance-ready — Supports GDPR auditing requirements

LyfPilot enforces hard boundaries between tenants at every level:

All 12+ third-party integrations follow the same security pattern:

1. OAuth2 with PKCE — No implicit flow, no client secrets in browser
2. Least-privilege scopes — Request only the permissions needed
3. Encrypted token storage — Tokens encrypted in vault, never in plaintext
4. Webhook signature verification — HMAC validation on incoming webhooks
5. Provider-specific error mapping — Graceful handling of revoked access

© 2024-2026 DevStudio AI Inc.. All rights reserved.