Autonomous SecOps Auto-Remediator

An enterprise-grade, autonomous Agentic AI system built to detect, audit, and remediate cloud security misconfigurations (such as public S3 buckets) in a completely isolated, simulated AWS environment.

This project demonstrates advanced AI system architecture, featuring multi-node state machines, persistent distributed memory, defense-in-depth guardrails, OpenTelemetry observability, and the Model Context Protocol (MCP).

---

Agentic Workflow Architecture

The agent operates as a LangGraph state machine. It utilizes a ReAct pattern to investigate the environment, drafts a remediation plan, validates it against strict guardrails, and pauses for human approval before executing securely via MCP.

graph TD
    %% Define styles
    classDef startend fill:#2d3748,stroke:#4a5568,stroke-width:2px,color:#fff,rx:20,ry:20;
    classDef llmNode fill:#2b6cb0,stroke:#2c5282,stroke-width:2px,color:#fff;
    classDef logicNode fill:#d69e2e,stroke:#b7791f,stroke-width:2px,color:#fff;
    classDef execNode fill:#c53030,stroke:#9b2c2c,stroke-width:2px,color:#fff;
    classDef db fill:#48bb78,stroke:#276749,stroke-width:2px,color:#fff;

    %% Nodes
    START((START)):::startend
    END_NODE((END)):::startend
    
    Gatekeeper{Incident<br>Already<br>Fixed?}:::logicNode
    Investigator[Investigator Node<br><i>Audit via AWS CLI</i>]:::llmNode
    Drafter[Drafter Node<br><i>Plan Remediation</i>]:::llmNode
    Guardrail[Guardrail Node<br><i>Pydantic Validation</i>]:::logicNode
    HITL{Human-in-the-Loop<br>Approval}:::logicNode
    Executor[Executor Node<br><i>FastMCP Shell</i>]:::execNode
    
    Mongo[(MongoDB Atlas<br>Checkpointer)]:::db

    %% Flow
    START --> Gatekeeper
    Gatekeeper -- Yes (Short Circuit) --> END_NODE
    Gatekeeper -- No --> Investigator
    
    Investigator --> Drafter
    Drafter --> Guardrail
    Guardrail --> HITL
    
    HITL -- Reject --> END_NODE
    HITL -- Approve --> Executor
    Executor --> END_NODE

    %% Memory interactions
    Investigator -. "Save State" .-> Mongo
    Drafter -. "Save State" .-> Mongo
    HITL -. "Freeze/Resume State" .-> Mongo

Loading

---

Tech Stack

• AI Orchestration: LangGraph (Stateful multi-agent workflows)

• LLM Engine: OpenRouter (gpt-4o-mini via LangChain)

• Tool Decoupling: Model Context Protocol (MCP) via FastMCP

• Cloud Infrastructure: LocalStack & Docker (Offline AWS simulation)

• Persistent Memory: MongoDB Atlas (Cross-session distributed checkpointer)

• Observability: Arize Phoenix & OpenTelemetry (Local LLMOps tracing)

---

Features

• Short-Circuit Memory Routing: The graph queries MongoDB Atlas at the entry node. If the thread_id indicates the incident is already remediated, the agent gracefully exits, preventing redundant LLM token usage.

• Model Context Protocol (MCP) Isolation: Tools are executed securely inside a containerized FastMCP server, completely isolating the AI's reasoning engine from the physical execution shell.

• Defense-in-Depth Guardrails: Pydantic validation strictly blocks destructive actions (e.g., delete-bucket) and bash-injection attempts before they can reach the execution layer.

• Human-in-the-Loop (HITL) Fault Tolerance: LangGraph checkpointers freeze the graph state directly to MongoDB prior to modifying cloud infrastructure. This allows the system to wait indefinitely for terminal approval, surviving reboots or crashes.

• OpenTelemetry Observability: fully instrumented with OpenInference. Traces are routed to a local Arize Phoenix server, providing X-ray visibility into graph transitions, tool latency, and token consumption without exposing data to third-party cloud trackers.

---

Getting Started

1. Prerequisites

• Python 3.10+

• Docker Desktop

• MongoDB Atlas cluster URI (free tier)

2. Environment Setup

Install the required dependencies:

pip install -r requirements.txt

Create a .env file in the root directory:

OPENROUTER_API_KEY="your-openrouter-key"
MONGODB_URI="mongodb+srv://<user>:<password>@cluster0.mongodb.net/?retryWrites=true&w=majority"

3. Start the Local Cloud

Spin up the LocalStack container. The infrastructure configuration automatically seeds a vulnerable S3 bucket (company-confidential-data) with a public-read ACL on boot:

cd infrastructure
docker-compose up -d

4. Run the Agent

python main.py

Recording-2.mp4

• Flow: Watch the agent investigate the environment, draft the fix, pause for your Y/N approval, and secure the bucket.

• Observability: While the agent is running, open http://localhost:6006 in your browser to view real-time OpenTelemetry traces in the Phoenix UI.

Recording-1.mp4