convez.ai
MCP ServerSegurançaAvançado
Bobcatsfan33

AI Security Platform

Bobcatsfan33/AI-Security-Platform

Control plane de segurança para IA com guardrails em runtime e remedição automática

Ver no GitHub
InicialPluginScore · de 0 a 100

Plataforma empresarial de segurança para IA com controle de runtime, red teaming automatizado e governança de compliance. Oferece avaliação de riscos (OWASP LLM Top 10), detecção de anomalias, agente inline em Go, conformidade regulatória (SOC 2, ISO 27001, EU AI Act) e integração com ferramentas SIEM/SOAR.

Instalação

$npx convez-ai Bobcatsfan33/AI-Security-Platform

1 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

Estrelas
2
Forks
1
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
v0.1.0
Issues abertas
23
Contribuidores
1

Casos de uso

  • Implementar guardrails de segurança em tempo real para modelos de linguagem
  • Executar red teaming automatizado e detectar vulnerabilidades em aplicações IA
  • Gerar relatórios de conformidade (SOC 2, FedRAMP, EU AI Act) para auditoria
  • Monitorar drift de modelos e risco de supply chain em IA-BOM
  • Integrar com SIEM/SOAR (Splunk, Datadog, PagerDuty) para resposta a incidentes

Veredito

Solução robusta e enterprise-grade para governança de IA, mas requer infraestrutura significativa (PostgreSQL, Redis, Redpanda) e expertise em DevOps. Ideal para organizações que precisam de compliance regulatório rigoroso.

README

A tradução para o português deste README ainda está na fila.

AI Security Platform

The control plane for enterprise AI security — evaluation, runtime protection, governance, and threat intelligence. Hybrid SaaS + on-prem product: this repository holds the multi-tenant control plane (Python / FastAPI), the customer-deployed runtime agent (Go), pluggable ONNX classifiers, a Next.js admin UI, and supporting SDKs for OpenAI / Anthropic.

Status: Tier 1 → Tier 3 engineering complete. Tier 4 (legal / marketing / sales artefacts) and a production hardening pass remain. See docs/ROADMAP.md for the original sprint sequence and docs/OPERATOR-RUNBOOK.md for day-2 ops.


Capabilities (what ships today)

SurfaceWhat it does
Evaluations50-case OWASP LLM Top 10 library + per-org cases; six model connectors (OpenAI, Anthropic, Ollama, Azure OpenAI, Bedrock, OpenAI-compat); LLM-judge + pattern verdicts
FindingsHash-chained audit trail through the open → in_progress → remediated → verified pipeline
Red teamGenerative campaigns with strategy library + judge; auto-promotion of successful attacks into the regression suite
AI-BOMAsset bill of materials, supply-chain risk scoring, model drift detection
Runtime agent (Go)Inline reverse proxy running all three policy stages live: Stage 1 regex/PII, Stage 2 ML (zero-config heuristic inline; ONNX inference sidecar via STAGE2_ONNX_ENDPOINT), Stage 3 LLM judge (deterministic default; configured judge via STAGE3_JUDGE_ENDPOINT, fail-open/closed per policy). Confidence-band routing, kill switch, telemetry to ClickHouse
SDKsPython + Node OpenAI/Anthropic wrappers that route through the local agent
ReportsMarkdown + PDF rendering of six templates (exec summary, technical detail, OWASP LLM Top 10, NIST AI RMF, SOC 2 AI, EU AI Act)
CI/CD gateComposite GitHub Action that triggers an evaluation, blocks the build on threshold breach, comments on PR
SIEM forwardersSplunk HEC, Elastic bulk, Sentinel HTTP Data Collector, Datadog Logs, Chronicle UDM, generic webhook
Dashboards/v1/dashboards/{runtime,traffic,policy-effectiveness} aggregations + Next.js executive view
Anomaly detectionPer-asset attack graph + statistical detector (volume spike / novel transition / risk inflation)
Threat intelOpt-in cross-tenant clustering; STIX 2.1 export
SOARPagerDuty / Opsgenie / generic-webhook incident sinks
ComplianceEvidence-pack ZIP for SOC 2, ISO 27001, FedRAMP Moderate

Architecture (binding decisions)

ConcernDecision
Operational dataPostgreSQL 16 + pgvector
TelemetryClickHouse (append-only)
Cache + pub/subRedis 7
Event streamingRedpanda (Kafka-compatible)
Control plane APIFastAPI (Python 3.12+)
Identity federationIDP-agnostic adapter; OIDC via authlib + joserfc, SAML via python3-saml, SCIM 2.0
Runtime agentGo 1.26 reverse proxy with httputil
ML classifierONNX models; Go-side runtime selected per-deployment
Policy enforcementThree-stage pipeline (regex / ML / LLM judge) per policy + per-org enforcement level
FrontendNext.js 16.2 App Router + Tailwind 4 + TypeScript

Running locally

# Generate a JWT secret
export JWT_SECRET=$(python -c "import secrets; print(secrets.token_urlsafe(64))")

# Bring up the stack
docker compose up -d postgres redis clickhouse redpanda
docker compose up app

API: http://localhost:8000/v1/docs. Frontend (in frontend/): npm run dev.

Apply migrations

cd backend
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env  # then edit JWT_SECRET
alembic upgrade head

Run tests

cd backend
pytest                              # full unit suite
pytest -m integration               # postgres + redis must be up
pytest --cov=app --cov-report=term-missing

499 unit tests pass on the current main.

Load test

pip install locust
locust -f backend/loadtest/locustfile.py --host http://localhost:8000 \
       -u 50 -r 5 --run-time 2m --csv loadtest_results

Repository layout

ai-security-platform/
├── backend/
│   ├── app/
│   │   ├── anomaly/              # Attack graph + detector
│   │   ├── api/v1/               # FastAPI routers
│   │   ├── auth/                 # JWT, API keys, RBAC
│   │   ├── compliance/           # Evidence-pack builder
│   │   ├── connectors/           # OpenAI/Anthropic/Bedrock/...
│   │   ├── db/                   # Models + Alembic
│   │   ├── evaluation/           # Runner
│   │   ├── identity/             # OIDC/SAML adapters
│   │   ├── policy/               # Three-stage pipeline
│   │   ├── redteam/              # Generative campaign engine
│   │   ├── reports/              # Markdown/PDF templates
│   │   ├── siem/                 # 6 export backends + forwarder
│   │   ├── soar/                 # 3 incident sinks
│   │   ├── telemetry/            # ClickHouse writer + dashboard queries
│   │   ├── threat_intel/         # Clustering + STIX export
│   │   └── main.py
│   ├── alembic/
│   ├── loadtest/locustfile.py
│   └── tests/{unit,integration}/
├── runtime-agent/                # Go 1.26 agent
├── frontend/                     # Next.js 16.2
├── sdks/{python,node}/           # Drop-in OpenAI/Anthropic wrappers
├── deploy/
│   ├── helm/ai-security-agent/
│   ├── k8s/agent.yaml
│   └── siem/{splunk,elastic,sentinel}/
├── actions/ai-security-gate/     # Composite GitHub Action
├── .github/workflows/ci.yml
└── docs/

License

BUSL-1.1 — Business Source License 1.1. Source-available; production use is permitted unless you are offering the work to third parties on a hosted or embedded basis to compete with the Licensor. Converts to Apache 2.0 four years after each version's publish date.

For alternative licensing, contact ryanwallac33@gmail.com.

Ver o original no GitHub

Bom para

Relacionados