Secure AI Agent Execution Platform 🛡️🤖

A system that allows AI agents to use tools autonomously while enforcing strict, enterprise-grade security policies and real-time defense monitoring.

🧨 The Problem

Modern AI agents that can interact with the outside world are susceptible to:

• Prompt Injection: An attacker overrides instructions to run malicious commands.
• Tool Abuse / LFI / SSRF: The agent queries unauthorized databases, internal domains, or sensitive file paths.
• Data Exfiltration: Sensitive user data is leaked through URL queries or remote hosts.

🛡️ The Solution (Architecture)

1. Policy Engine: JSON-based rule system enforcing domain allowlists, accepted HTTP methods, and read-only DB permissions.
2. Sandbox Layer (Executor): A sterile execution wrapper preventing direct API calls and filtering requests through the Policy Engine.
3. Defense Monitor: Real-time state tracking and anomaly detection. Catches Rate-limit abuses, suspicious SQL payloads, and URL exfiltration signatures before execution.
4. Adversarial Attack Simulator: A comprehensive script acting as the "Attacker", validating that the aforementioned defenses hold up against Prompt Injection, SQLi, and DDOS.

🛠️ Components

• agent/: Simple simulated LLM routing intents to tools.
• tools/: Standardized BaseTool interfaces (HTTP, DB, File).
• policy/: Configuration (policies.json) and the PolicyValidator.
• sandbox/: The secure wrapper (SecureToolExecutor).
• defense/: AnomalyDetector, logging, and immediate response triggers.
• evaluation/: Generates metrics reports on allowed vs blocked requests.
• attacks/: AttackRunner that hits the system with malicious payloads to verify robustness.

🚀 Usage

Run the standard benign Agent:

python3 main.py

Run the Adversarial Attack Simulator:

python3 attacks/runner.py

This will run SQL injection attempts, an Exfiltration attempt via URL, and an internal server SSRF attempt, proving that the Sandboxing and Monitoring actively drop the requests.