tyk

Least-privilege permissions

• Limit credential access: Only give the gateway credentials for the specific downstream services it needs to reach. If a backend service gets compromised, attackers won't automatically gain access to all your other systems.
• Restrict agent capabilities: Configure what each AI agent is allowed to do through the gateway (which APIs it can call, what data it can read/write). Block unnecessary access paths.
• Separate sensitive endpoints: Keep high-risk services (payment systems, HR databases, customer records) behind stricter access rules than public-facing APIs.
• Rotate credentials regularly: Change passwords and API keys for services connected through the gateway every 90 days or when staff leave.

Network isolation

• Use a security boundary: Place the gateway on its own network segment (DMZ) so if it's compromised, attackers can't immediately pivot into your internal systems.
• Restrict outbound connections: Configure firewall rules so the gateway can only reach the specific internal services and external APIs it legitimately needs—block all other traffic.
• Require encrypted connections: Enable TLS/HTTPS for all communication flowing through the gateway (both incoming agent requests and outgoing service calls).
• Limit who can access the gateway: Use IP whitelisting or VPN to ensure only authorized applications and users can connect to the gateway itself.

Monitoring

• Log all requests and responses: Record what data moves through the gateway, who requested it, and what the gateway returned. Store logs securely and keep them for at least 90 days.
• Alert on suspicious patterns: Set up automatic warnings if someone tries unusual access patterns (same user accessing many different services rapidly, large data downloads, failed login attempts).
• Track configuration changes: Maintain an audit trail of when and how gateway settings, credentials, or access rules are modified.
• Monitor for unusual behavior: Watch for signs the gateway is being used to exfiltrate data (unexpected large transfers) or inject malicious instructions into AI agents.

Red flags to watch for

• Unpatched versions in production: If your open-source gateway version is several months old and security updates are available, update immediately—delays put you at known risk.
• Overly broad access rules: If the gateway allows agents to access any service or if credentials aren't scoped to specific systems, you have a single-point-of-failure vulnerability.
• Disabled logging or validation: If request validation is turned off for performance, or if logs are deleted quickly to save storage, you've lost visibility into attacks and compliance evidence.
• No separation between test and production: If your staging environment uses the same credentials and network access as production, a developer mistake or test data breach can become a real incident.