Skip to main content

Security Model

How UAPK Gateway provides security for AI agent operations.

Authentication​

Human Operators​

  • JWT tokens issued after login
  • Configurable expiration
  • Role-based access control (future)

AI Agents​

  • API keys for identification
  • Capability tokens for authorization
  • Tokens are scoped and time-limited

Authorization​

Capability-Based Security​

Agents can only perform actions explicitly granted:

  1. Agent registers with a Manifest
  2. Operator reviews and approves capabilities
  3. Gateway issues Capability Token
  4. Every request is checked against token

Policy Enforcement​

Multiple layers of policy checks:

Request → Auth → Capability Check → Policy Engine → Execute
                                          ↓
                                    Rate Limits
                                    Budget Limits
                                    Time Restrictions
                                    Custom Rules

Audit Trail​

Tamper-Evident Logging​

Every InteractionRecord includes:

  • Signature: HMAC of record contents
  • Chain Hash: Hash including previous record

This creates a blockchain-like chain where any modification is detectable.

What's Logged​

  • All action requests (approved or denied)
  • Policy evaluation results
  • Execution results or errors
  • Timestamps and agent identification

Data Protection​

Secrets Management​

  • Secrets via environment variables
  • Never logged or exposed in API
  • Database credentials isolated

Network Security​

  • TLS everywhere in production
  • Internal services on Docker network
  • No direct database access from outside

Threat Model​

Threats Addressed​

ThreatMitigation
Agent exceeds scopeCapability tokens + policies
Unauthorized actionMulti-layer auth checks
Audit log tamperingSigned, chained records
Credential theftShort-lived tokens, revocation
Resource exhaustionRate limits, budgets

Out of Scope (v0.1)​

  • Multi-region deployment
  • Hardware security modules
  • Zero-knowledge proofs

Best Practices​

For Operators​

  1. Use strong, unique SECRET_KEY
  2. Review agent manifests carefully
  3. Start with restrictive policies
  4. Monitor audit logs regularly
  5. Rotate capability tokens

For Agent Developers​

  1. Request minimum necessary capabilities
  2. Include meaningful context in requests
  3. Handle policy denials gracefully
  4. Don't store capability tokens long-term