Skip to content

☁️ 🤖 LLM agent-based simulations to generate benign and malicious Cloud logs

License

Notifications You must be signed in to change notification settings

TracecatHQ/sims

Repository files navigation

☁️🤖 Sims

Next.js FastAPI OpenAI Discord

Run LLM agent-based simulations to generate benign and malicious Cloud logs. Simulate multi-chain attacks using a click-and-drag attack builder. Try it live here: https://simulation.tracecat.com/workbench

We've only implemented AWS attack techniques. But sims can be easily extended to include other Cloud platforms. If you are interested in using or improving this project, please join our Discord for Q&A and updates!

Screenshot

Motivation

We developed AI Agents from scratch to simulate both malicious and benign behavior in the Cloud. This library relies on Stratus Red Team, an open-source advesary emulation library from DataDog, to define individual attack techniques.

Existing advesary emulation labs generate examples of true positive alerts. These are great for writing detections for known malicious behavior, but these labs don't provide examples of false positives: i.e. why a non-malicious user (e.g. software engineer) might use one of the procedures in the attack in their day-to-day job.

Discloure: the results are still worse than what you would get from a manually built lab in a live AWS environment. The benefit from this AI agents approach, however, is the ability to generate stories to better understand false positives.

LLM Workflow

We have two independent AI agents, MaliciousStratusUser and NoisyStratusUser, that run async. These agents are used to generate true positive and false positive behavior associated with a specific attack technique.

Note: Agent "thoughts" and synthetic Cloud logs are saved as ndjson in ~/.sims/lab/thoughts.

LLM Workflow

Deployment

API

To run the simulations yourself, spin up the sims FastAPI server and call the /labs/ws websocket endpoint.

Local

Please set the env variables in .env.local before running commands.

cp .env.local.example .env.local
# .env.local

# Use 'dev' for local development
TRACECAT__ENV=dev

# OpenAI API key
OPENAI_API_KEY=...

# Optional: OpenAI organization ID
OPENAI_ORG_ID=...

Deploy the FastAPI app using uvicorn. You may wish to specify the number of workers with the --workers flag.

uvicorn sims.api.server:app --reload

Modal Cloud

We use Modal for serverless deployments. Please set the env variables in .env.modal before running commands.

cp .env.modal.example .env.modal
# .env.modal

# 'dev' or 'prod'.
# Use 'dev' when serving the Modal endpoint from the CLI.
# Use 'prod' when deploying to Modal.
TRACECAT__ENV=prod

# The name of the secret you created in the Modal dashboard
TRACECAT__MODAL_OPENAI_SECRET_NAME=...

# The URL of the client
TRACECAT__FRONTEND_URL=...

You will also have to setup a Modal account and the CLI tool to proceed. Please follow the instructions here.

Serving the Modal endpoint with hot reload (for development):

modal serve sims/api/modal_server.py

For full deployment:

modal deploy --name <deployment name> sims/api/modal_server.py

Frontend

Local

Follow the pnpm installation instructions here.

Configure your frontend/.env.local file to point to the API endpoint URL using NEXT_PUBLIC_API_URL.

# frontend/.env.local

# Your API endpoint URL
NEXT_PUBLIC_API_URL=http://localhost:8000

Then run the development server with pnpm:

cd frontend
pnpm dev

Vercel

We recommend depoying the frontend on Vercel, override the NEXT_PUBLIC_API_URL in the Vercel dashboard and point the Root Directory to frontend.

About

☁️ 🤖 LLM agent-based simulations to generate benign and malicious Cloud logs

Resources

License

Stars

Watchers

Forks

Contributors 3

  •  
  •  
  •