AgentGate

Open-Source AI Agent Security Monitor with MCP Security Gateway

The Problem

AI agents are the fastest-growing attack surface in enterprise software:

• 1 in 8 AI breaches are linked to agentic systems (HiddenLayer 2026)
• 76% of organizations cite shadow AI as a top risk
• Only 6% have advanced AI security controls
• CVE-2025-6514 (CVSS 9.6) demonstrated critical MCP vulnerabilities
• The Clinejection attack (Feb 2026) turned an AI coding bot into a supply chain vector affecting 5M+ users
• Zero open-source tooling comprehensively addresses agentic AI runtime security

AgentGate fills this gap.

Features

• MCP Security Gateway — Intercepts and enforces policies on all MCP tool calls
• Attack Detection — Prompt injection, data exfiltration, privilege escalation, tool abuse
• Behavioral Anomaly Detection — ML-based (Isolation Forest) behavioral baseline monitoring
• RAG Poisoning Scanner — Detects poisoned documents in knowledge bases
• Shadow AI Discovery — Identifies unauthorized agents and MCP connections
• Transparent Risk Scoring — Documented algorithm with full methodology
• YAML Policy Engine — Configurable policies mapped to OWASP & MITRE frameworks
• AI Intelligence — LangChain-powered summaries, recommendations, NL queries
• Real-time Dashboard — Dark/light mode React UI with live feeds and analytics
• Compliance Reports — PDF reports for OWASP Agentic Top 10 & MITRE ATLAS
• Live Demo Agent — Real LangChain agent routed through the security gateway
• DevSecOps Pipeline — Bandit, Trivy, Gitleaks, pip-audit, OWASP ZAP scanning

Architecture

graph LR
    A[AI Agent] -->|MCP| B[Gateway]
    B --> C{Analysis Engine}
    C --> D[Attack Detector]
    C --> E[Behavioral Analyzer]
    C --> F[RAG Scanner]
    C --> G[Shadow AI]
    D --> H[Policy Engine]
    E --> H
    H -->|ALLOW| I[MCP Server]
    H -->|BLOCK| A
    H --> J[Dashboard]
    H --> K[Alerts]

Loading

Tech Stack

Layer	Technology
Backend	Python 3.11, FastAPI, SQLAlchemy, Pydantic v2
ML/AI	scikit-learn, LangChain, OpenRouter
Frontend	React 18, Tailwind CSS, Recharts, Lucide React
Database	PostgreSQL 16 (prod) / SQLite (dev)
Deployment	Docker, Docker Compose
CI/CD	GitHub Actions
Security	Bandit, Trivy, Gitleaks, pip-audit, OWASP ZAP

Quick Start

# Clone and start
git clone https://github.com/yourusername/agentgate.git
cd agentgate
docker-compose up --build

# Open dashboard
open http://localhost:3000

# Seed demo data
docker exec -it agentgate-backend-1 python seed_data.py

Environment Variables

Variable	Default	Description
DATABASE_URL	SQLite	Database connection string
OPENROUTER_API_KEY	-	OpenRouter API key for AI features
OPENROUTER_MODEL	openai/gpt-4o-mini-2024-07-18	LLM model
API_KEY	agentgate-demo-key-2026	API authentication key
SLACK_WEBHOOK_URL	-	Slack alert webhook
TEAMS_WEBHOOK_URL	-	Teams alert webhook
ALERT_ENABLED	false	Enable alert delivery

OWASP Agentic Top 10 Coverage

Category	Detection	Policy
A01 - Prompt Injection	Pattern + ML	BLOCK at 0.7 confidence
A02 - Sensitive Data Exposure	PII/credential regex	BLOCK
A03 - Sandboxing Failures	Tool abuse patterns	ALERT
A04 - Unauthorized Code Execution	Privilege escalation	BLOCK at 0.7
A05 - SSRF Vulnerabilities	URL pattern analysis	ALERT
A06 - Excessive Permissions	Unapproved server blocking	BLOCK
A07 - Data Leakage	Behavioral anomaly detection	ALERT
A08 - Insufficient Logging	Full audit trail	Always logged
A09 - Overreliance	Agent quarantine	BLOCK quarantined
A10 - Model Theft	Access monitoring	ALERT

MITRE ATLAS Coverage

Technique	ID	Detection
Prompt Injection via MCP	AML.T0051	Pattern + ML classifier
Data Exfiltration via Tools	AML.T0048	PII/credential scanning
Privilege Escalation	AML.T0052	Sensitive path + code detection
Tool Shadowing	AML.T0054	Levenshtein + description analysis

Live Demo

See DEMO.md for instructions.

Mode 1 — Simulated Traffic (2 min): Dashboard with seed data + curl attacks

Mode 2 — Live Agent Interception (3 min): Real LangChain agent routed through AgentGate

# Run the live demo
bash demo/run_demo.sh

Documentation

• ARCHITECTURE.md — System architecture
• API_DOCS.md — API reference
• RISK_SCORING.md — Risk scoring methodology
• THREAT_MODEL.md — Threat model
• DEMO.md — Demo instructions

Future Roadmap

• Kubernetes Helm charts
• SIEM integration (syslog/webhook)
• Browser extension for real-time monitoring
• Slack bot for interactive investigation
• Multi-agent orchestration monitoring
• Agent-to-agent communication monitoring

Contributing

Contributions welcome! Please open an issue or PR.

License

MIT License