AI Codebase Audit Engine
ojackson08/ai-codebase-audit-engine
Auditorias de segurança de código automatizadas com IA em minutos
Ver no GitHubPipeline serverless orientado a eventos que automatiza auditorias de segurança de código usando Claude 3 via Amazon Bedrock. Acionado por upload em S3, chunifica e analisa bases de código, gerando relatórios estruturados de vulnerabilidades e tarefas remediáveis em minutos.
Instalação
npx convez-ai ojackson08/ai-codebase-audit-engine1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Análise automatizada de segurança em grandes bases de código
- Detecção de vulnerabilidades de injeção de prompt e padrões inseguros
- Auditoria de riscos em supply chain de dependências
- Geração de tarefas paralelizáveis para remediação automática
Veredito

Solução robusta e pronta para produção que reduz drasticamente o tempo de auditoria de segurança. Bem arquitetada com Terraform/AWS, ideal para equipes que auditam aplicações de IA, mas exige familiaridade com infraestrutura cloud e Claude via Bedrock.
README
A tradução para o português deste README ainda está na fila.
AI Codebase Audit Engine
GenAI-powered security code review pipeline — the core engine behind Merkaba AI Risk Management's AI Codebase Security Audit service.
Overview
The AI Codebase Audit Engine is an event-driven GenAI pipeline that transforms weeks of manual security code review into minutes of automated, AI-assisted analysis. A codebase upload to S3 triggers a Lambda function that chunks and embeds the code, queries Amazon Bedrock (Claude 3), and produces a structured, parallelizable set of security findings and remediation tasks.
This tool is the backbone of Merkaba AI Risk Management's AI Codebase Security Audits service — used to identify insecure patterns, prompt injection vulnerabilities, supply chain risks, and architectural security flaws in AI-powered applications.
Architecture
S3 Upload (codebase .zip)
│
▼
Lambda (S3 event trigger)
│
├── Extract + chunk code files
│
├── Query Amazon Bedrock (Claude 3 Sonnet)
│ └── Security analysis per chunk
│
├── Store report → DynamoDB
│
└── Generate structured output
├── findings.json
└── agent_tasks.md (parallelizable remediation tasks)
Capabilities
| Capability | Description |
|---|---|
| S3-triggered analysis | Upload any codebase as a ZIP; Lambda handles the rest automatically |
| Intelligent chunking | Splits large codebases into semantically meaningful units for analysis |
| Claude 3 / Bedrock integration | Leverages frontier LLMs for nuanced security reasoning |
| Insecure pattern detection | Identifies hardcoded secrets, SQL injection, insecure deserialization, and more |
| Prompt injection risk analysis | Flags code paths where user input reaches LLM prompts unsanitized |
| Supply chain analysis | Reviews dependency usage for known vulnerable patterns |
| Parallelizable output | Generates agent-ready task lists for downstream remediation automation |
Infrastructure
Provisioned entirely via Terraform:
- AWS Lambda — Serverless analysis engine
- Amazon S3 — Codebase ingestion and report storage
- Amazon Bedrock — Claude 3 Sonnet inference
- Amazon DynamoDB — Audit report and agent task storage
- IAM — Least-privilege execution roles
Deployment
# Provision infrastructure
cd terraform/
terraform init
terraform apply
# Upload a codebase for analysis
aws s3 cp ./my-project.zip s3://your-audit-bucket/uploads/
# Retrieve the security report
aws s3 cp s3://your-audit-bucket/reports/latest/ ./reports/ --recursive
Case Study / Usage Notes
Deployment at Merkaba AI Risk Management:
This engine was deployed during a security audit engagement for a Series A startup building an AI-powered legal document analysis platform. The codebase (~18,000 lines of Python) was uploaded to the S3 trigger bucket. Within 4 minutes, the engine produced a 47-item security findings report covering: 3 hardcoded API keys, 11 instances of unsanitized user input reaching LLM prompts, 2 insecure deserialization patterns, and a critical supply chain risk in a pinned dependency. The parallelizable task output was fed directly into the client's Jira backlog, reducing triage time from an estimated 3 days to under 2 hours.
Integration with Merkaba Security Stack
merka-prompt-shield— Deploy after audit to protect identified injection-vulnerable endpointsagent-security-scanner— Companion tool for agent configuration analysisaws-security-compliance-automation— Infrastructure hardening after codebase auditredteam-kit— Adversarial validation of findings
License
MIT License — see LICENSE for details.
Contact
Merkaba AI Risk Management security@merkabacreatives.org https://merkabacreatives.org/ai-risk Atlanta, GA — Remote Worldwide