CFAM: Middleware de Filtragem para IA 🇧🇷

O que é

O CFAM é uma ferramenta de linha de comando (CLI) que funciona como uma camada de segurança entre os modelos de IA e o seu sistema interno. Ele garante que só passe o que é seguro e bem formatado.

Modelos de IA geram respostas imprevisíveis. O CFAM resolve isso aplicando filtros rígidos e determinísticos.

O que ele faz

• Intercepta o Dado: Captura a resposta da IA imediatamente após ela ser gerada.
• Aplica Filtros Determinísticos: Avalia o conteúdo com lógica rígida. Se não estiver no formato exato esperado, é rejeitado.
• Impede Travamentos: Funciona como um fusível. Se a IA mandar um dado quebrado, o CFAM bloqueia e seu sistema não cai.
• Gera Log de Auditoria: Grava cada decisão (aceito ou bloqueado) no arquivo cfam_audit_ledger.log.

Soluções de Segurança

• Bloqueio de Payload Malicioso: Intercepta e descarta injeções de código e tentativas de roubo de dados.
• Proteção contra Crash: Protege suas APIs e bancos de dados de dados malformados.
• Superfície de Ataque Reduzida: É um binário compilado único, sem depender de um monte de pacotes externos.
• Rastreabilidade: Tudo fica registrado localmente para análise forense.

Status: Prova de Conceito (PoC)

Este software é atualmente uma Prova de Conceito experimental, para fins de pesquisa e testes. Prioriza a integridade do dado.

Como Funciona o Filtro

1. Integridade Estrutural: Bloqueia payloads cortados ou com formatação quebrada.
2. Limite de Fronteira: Rejeita payloads muito grandes ou com caracteres inválidos.
3. Rejeição de Anomalia: Bloqueia estruturas de ataque conhecidas.

Exemplo

• Aceito: Payload completo e bem formado. (Retorna Exit Code: 0)
• Rejeitado: Resposta cortada ou malformada. (Bloqueia e retorna Exit Code: 1)

Decisões de Arquitetura

• Feito em Rust: Para garantir segurança de memória, baixa latência e baixo consumo.
• Zero Dependências: Roda isolado, sem precisar de interpretador.

--

CFAM: Core Filtering AI Middleware 🇺🇸

Overview

CFAM is a command-line interface (CLI) tool that acts as a deterministic filtering layer between AI models and your internal systems. Its primary function is to enforce structural and logical rules on AI outputs.

What It Does

AI models generate probabilistic and unpredictable responses. CFAM solves this by applying strict, rule-based validation to ensure only safe, structured, and compliant data passes through to your applications.

• Intercepts Data: Captures the AI payload immediately after generation.
• Applies Deterministic Filters: Evaluates the payload using rigid, non-probabilistic logic. If the data does not match the exact expected format or rules, it is rejected.
• Enforces Governance: Acts as a hard-coded barrier to prevent system crashes or logic errors caused by AI hallucinations.
• Generates Audit Logs: Records every decision (accepted or blocked payloads) in a local ledger file for full operational transparency.

• Security Solutions

CFAM provides a hard security layer for AI integrations by addressing vulnerabilities that traditional software frameworks leave exposed:

• Malicious Payload Blocking: Intercepts and discards AI outputs that contain injection attacks, malicious code, or unauthorized data extractions before they reach your internal infrastructure.
• System Crash Protection: Acts as a technical fuse. If an AI model returns broken, incomplete, or malformed data, CFAM blocks the transaction immediately, protecting downstream APIs and databases from crashing.
• Reduced Attack Surface: Built as a standalone, compiled binary. It eliminates the security risks associated with large trees of third-party software dependencies common in interpreted environments.
• Traceability and Audit Ledger: Records every single filtering decision into a permanent, local log file. This enables immediate forensic analysis of blocked payloads and model compliance.

Status: Proof of Concept (PoC)

This software is currently an experimental Proof of Concept (PoC). It is designed strictly for research, testing, and evaluation purposes. At this stage, the system prioritizes logical enforcement and data integrity over high-scale distributed performance.

Verification

To ensure the integrity of the downloaded binary, verify its SHA-256 checksum against the values provided in the official GitHub Releases page.

Filter Contract & Validation Rules

CFAM inspects incoming string payloads via strict structural analysis. It evaluates data against three non-probabilistic criteria:

1. Structural Integrity: Detects and blocks truncated payloads or broken formatting (e.g., incomplete strings or unclosed markers).
2. Boundary Enforcement: Rejects payloads that exceed structural length constraints or contain invalid control characters.
3. Anomaly Rejection: Blocks payloads containing known adversarial structures or system-level escape sequences.

Example

• Accepted Input: A complete, well-formed payload that strictly respects formatting boundaries. (System returns Exit Code: 0).

• Rejected Input: Truncated AI responses, unclosed delimiters, or malformed data streams. (System blocks execution and returns Exit Code: 1).

  Architectural Decisions

• Development in Rust: Built strictly for memory safety, low latency, and predictable resource allocation.

• Zero Runtime Dependencies: Distributed as a statically compiled binary. It runs entirely isolated from interpreters, virtual machines, or external package managers, reducing the system's attack surface.

Usage

CFAM operates as a standalone binary executable. cfam_audit_ledger.log

./cfam "Your input or task description"

Author & Support

Developed by Marta Reinhardt For verification, security inquiries, or bug reports regarding the compiled binary, please use the official GitHub repository issue tracker.

Copyright © 2026 Marta

All Rights Reserved.