AI features run inside the same security boundary as the rest of the Gabaltech Platform. There is no separate, more-trusted path for AI-driven actions.
Prompt security
User input passed to AI models is treated as untrusted input, the same way we treat any external data. Prompts are not used to construct shell commands, SQL queries or infrastructure changes without going through normal validation.
Secure infrastructure
AI features run on the same ECS Fargate infrastructure, VPC boundaries and security groups as everything else — no bespoke, less-audited AI infrastructure sitting outside our normal controls.
Least privilege
AI-assisted workflows use scoped IAM roles limited to exactly what that workflow needs. A code-review assistant doesn't get infrastructure-deploy permissions; a deployment assistant doesn't get billing access.
Secrets protection
AI never sees raw secrets. Credentials are injected at runtime via SSM Parameter Store / Secrets Manager, never hardcoded, and never included in prompts sent to a model.
No hardcoded credentials
This is a hard rule across the platform, enforced by secret-scanning in CI — AI-generated code is scanned the same way human-written code is.
Security scanning
AI-generated code and infrastructure changes pass through the same static analysis, dependency scanning and policy checks as any other change before merge.
Infrastructure validation
AI-generated Terraform and Pulumi changes are validated (plan, preview) before they can be applied — nothing is applied blind.
Audit logging
Every AI-assisted action that touches infrastructure or deployments is logged with who requested it, what was generated, and what was actually applied.