Cloud Attack Range

A self-hosted, deliberately vulnerable cloud lab for authorized security training and AI-agent benchmarking. Run the interactive Python entrypoint on macOS, Linux, Windows, or WSL to create the lab in an AWS, Azure, or GCP scope that you own.

Caution

This project creates weak identities, exposed synthetic data, mutable workloads, and optionally destructive permissions. Use a dedicated, disposable, non-production cloud account. Your workstation and Terraform state must stay outside the lab. Nothing cleans up automatically. You run python3 cloudrange.py --teardown RANGE_ID yourself. Read Safety and isolation before setup.

New here? Start with the operator guide.

What you get

• A realistic lab. Identities, networks, storage, secrets, compute, serverless, messaging, and registries that interact instead of isolated, one-off simulations.
• 239 catalogued weaknesses, each with a permanent BAR-* ID, grouped into 44 categories. See what is vulnerable in the lab.
• A benchmark harness. Run repeatable agent evaluations and score the results offline.
• Safe defaults. Public exposure, destructive paths, costly services, host access, and tenant-wide paths are each a separate opt-in.

The lifecycle

flowchart LR
    A[Clone repository] --> B[Log in to provider]
    B --> C["setup<br/>builds control plane"]
    C --> D["plan<br/>saves a bound plan"]
    D --> E["deploy<br/>builds the weak lab"]
    E --> F[Verify and test]
    F --> G["teardown<br/>you run this"]
    G --> H["verify-clean<br/>proves it is gone"]

Loading

setup builds the control plane; deploy builds the vulnerable lab.

Install

Python 3.11 or newer is required. Install Python, Git, Terraform 1.8+, and the CLI for your provider (aws, az, or gcloud).

With Claude Code or Codex

Open either coding agent and paste this prompt:

Set up Cloud Attack Range from https://github.com/bugbasesecurity/cloud-attack-range on this machine. Read README.md and docs/wiki/setup.md. Ask me the questions needed to choose the configuration, then follow the documented workflow. Stop before deployment so I can review the plan.

Manually

git clone https://github.com/bugbasesecurity/cloud-attack-range.git
cd cloud-attack-range
python3 cloudrange.py --help

Keep this checkout in place until teardown and verify-clean finish. On native Windows, run py -3 cloudrange.py --setup. If preflight reports missing tools, run python3 cloudrange.py --install-deps --provider aws, replacing aws with your provider.

Run a lab

Set a provider budget alert before you start.

python3 cloudrange.py --preflight --provider aws      # read-only check of this machine
aws login --profile range-bootstrap       # or: az login / gcloud auth login
python3 cloudrange.py --setup                         # wizard; builds the control plane
python3 cloudrange.py --doctor RANGE_ID               # readiness report
python3 cloudrange.py --plan RANGE_ID                 # saved, digest-bound plan
python3 cloudrange.py --deploy RANGE_ID               # creates the vulnerable lab
python3 cloudrange.py --status RANGE_ID

Azure setup cannot create the disposable Entra tenant or its first subscription. Create both manually. On a remote machine, select them with:

az login --use-device-code --tenant YOUR_DISPOSABLE_TENANT_ID
az account set --subscription YOUR_DISPOSABLE_SUBSCRIPTION_ID

See Azure bootstrap for the required IDs and roles.

When you are done, run:

python3 cloudrange.py --teardown RANGE_ID
python3 cloudrange.py --verify-clean RANGE_ID

Full walkthrough: Setup from zero.

Lab presets

Preset	Cloud layout	Use it for
Core Safe	One AWS, Azure, or GCP scope	First run, single-cloud benchmarking
Containers	Core plus EKS, AKS, or GKE and attack pods	Kubernetes and workload identity
Advanced	One provider, individually selected controls	Broader reviewed campaigns

Advanced does not enable everything; each control is confirmed separately.

Range ID

A Range ID is the permanent name of a lab instance, such as research-aws-01. Setup asks you to choose one, and lifecycle commands use it to identify the range. It cannot be renamed or reused. See the Range ID contract.

Benchmarking

The benchmark harness provides a provider-neutral run format and offline scorer. The solutions are published in what is vulnerable in the lab, so keep evaluated models away from that page and this repository. See the benchmark protocol.

Documentation

• Operator guide
• Setup from zero · Troubleshooting
• Command reference · Provider guides
• Safety · Costs

License

Apache License 2.0. The license does not authorize testing systems you do not own.