Bedrock-Guardrail-Firewall

Overview

Bedrock Guardrail Firewall is a privacy-first enforcement layer for generative AI systems. It combines deterministic local controls, optional Microsoft Presidio analysis, and the Amazon Bedrock ApplyGuardrail API without invoking a foundation model. The production runtime remains a single Python file: orchestrator.py . Policy, documentation, tests, and repository automation are kept separate so the runtime stays portable and auditable. The system evaluates user input, candidate output, and retrieval context. It can sanitize detected sensitive data before optional Bedrock Guardrails calls, and a local review, escalation, or block decision prevents live content transmission. After optional cloud processing, Bedrock-transformed content is size-checked and re-evaluated by local privacy, policy, and grounding controls. The runtime returns one of five actions: allow, sanitize, queue_for_review, escalate, or block. Sanitized content is released only for allow and sanitize recommendations. AWS safety modes include disabled, preview, and live. The CLI requires both --aws-mode live and --enable-live-aws . Policy profiles include balanced, production, and offline_test. The project is intended as a policy enforcement point around application traffic, not a network firewall, web application firewall, IAM system, malware scanner, factuality oracle, or replacement for human approval. A caller must honor the returned action, content-release flag, and capability restrictions.