loader/_docs/00_problem/security_approach.md

Security Approach

Authentication

• Mechanism: JWT Bearer tokens issued by Azaion Resource API
• Token handling: Decoded without signature verification (options={"verify_signature": False}) — trusts the API server
• Token refresh: Automatic re-login on 401/403 responses (single retry)
• Credential storage: In-memory only (Credentials object); not persisted to disk

Authorization

• Model: Role-based (RoleEnum with 7 levels: NONE through ApiAdmin)
• Enforcement: Roles are parsed from JWT and stored on the User object, but no endpoint-level authorization is enforced by the loader. All endpoints are accessible once credentials are set.

Encryption

Resource Encryption (binary-split scheme)

• Algorithm: AES-256-CBC with PKCS7 padding
• Key expansion: SHA-256 hash of string key → 32-byte AES key
• IV: Random 16-byte IV prepended to ciphertext

Key Derivation

Key Type	Derivation	Scope
API download key	SHA-384(email + password + hw_hash + salt)	Per-user, per-machine
Hardware hash	SHA-384("Azaion_" + hardware_fingerprint + salt)	Per-machine
Resource encryption key	SHA-384(fixed_salt_string)	Global (shared across all users)
Archive decryption key	SHA-256(key_fragment_from_api)	Per-unlock operation

Binary Split

• Resources encrypted with shared resource key, then split into:
  • Small part (≤3KB or 30%): uploaded to authenticated API
  • Big part (remainder): uploaded to CDN
• Decryption requires both parts — compromise of either storage alone is insufficient

Hardware Binding

• Hardware fingerprint: CPU model, GPU, memory size, drive serial number
• Used to derive per-machine encryption keys for API resource downloads
• Prevents extraction of downloaded resources to different hardware

IP Protection

• Security-sensitive modules (security, api_client, credentials, etc.) are Cython .pyx files compiled to native .so extensions
• Key derivation salts and logic are in compiled code, not readable Python

Secrets Management

• CDN credentials stored in cdn.yaml, downloaded encrypted from the API
• User credentials exist only in memory
• JWT tokens exist only in memory
• No .env file or secrets manager — environment variables for runtime config

Input Validation

• Pydantic models validate request structure (LoginRequest, LoadRequest)
• No additional input sanitization beyond Pydantic type checking
• No rate limiting on any endpoint

Known Security Gaps

1. JWT decoded without signature verification
2. No endpoint-level authorization enforcement
3. No rate limiting
4. Resource encryption key is static/shared — not per-user
5. subprocess with shell=True in hardware_service (not user-input-driven, but still a risk pattern)
6. No HTTPS termination within the service (assumes reverse proxy or direct Docker network)