Cyberful: The Open-Source AI Red Team.

For authorized pentests, code audits, and bug bounty research.

Cyberful is an AI Red Team. In other words, it is your security research laboratory that uses a multi-agent system to orchestrate 316 cybersecurity tools, including 203 command-line applications and commands such as Nmap, Metasploit, and Nuclei, alongside ZAP and Ghidra.

In short, hypothesis workers map attack surfaces, test exploit paths, and validate impact in isolation. Then, a deterministic control plane enforces scope and preserves evidence. Finally, only independently reproduced vulnerabilities reach reports. This applies to penetration testing, code audits, and bug bounty hunting.

(i) Documentation: installation, workflows, configuration, architecture, and security runtimes are documented at cyberful.ai/open-docs.

Why Cyberful

Cyberful exists to democratize cybersecurity: to make advanced security work understandable, inspectable, and usable by more developers, researchers, maintainers, and smaller teams. Democratization does not mean unrestricted automation or lower standards; findings remain evidence-backed, independently verifiable, and tied to the engagement that authorized them.

Getting started

Cyberful requires Node.js 18 or newer, npm, Docker, and a configured model provider.

1. Install Cyberful.

npm install --global cyberful

2. Choose or create a folder that will serve as your cybersecurity laboratory.

mkdir cyberful-lab
cd cyberful-lab

3. Create a settings.yaml based on example.settings.yaml, then authenticate the configured provider.

cyberful auth login

4. Start Cyberful and select a workflow.

cyberful

For complete fresh-machine instructions on macOS, Linux, and Windows, follow Your first penetration test.

To keep model inference on your own machine, follow Local inference.

Requirements

To install through npm and run a release:

• Docker with a running Linux-container engine;
• Node.js 18 or newer with npm;
• one configured model provider;
• at least 100 GB of free disk space before the first runtime build;
• at least 10 GB of RAM dedicated to Docker.

ZAP, Ghidra, Python, Ruby/Bundler, native debugging and fuzzing, managed Firefox/Marionette, Xvfb/X11 clipboard testing, archive extraction, and the offensive toolchain are included in the runtime image. Before a live-target AgentRun starts, an ephemeral private-network HTTPS canary verifies that curl/OpenSSL, Git, Requests/pip, Node, and Ruby/Bundler can traverse the real ZAP proxy with the attested engagement CA; it never contacts the target or Internet. Brief then installs and attests the engagement's host-scoped rate limit and mandatory public request headers in ZAP before numbered target-profile preflight. After each accepted Pentest or Bug Bounty phase, the host records a best-effort passive ZAP checkpoint filtered to observed origins whose hostname is authorized by the finalized policy; non-web engagements make no checkpoint ZAP call or wait, and collection never blocks handoff. Cyberful embeds its pinned hardened agent-browser fork and prepares an isolated compatible Chrome for agent-controlled browsing, so it never needs access to a personal browser profile. Docker Compose is not required.

See What you need for the supported host platforms, release architectures, provider setup, and source development requirements.

Build and test

From the repository root:

make deps       # install workspace and MCP dependencies
./scripts/update_pi.py  # update and attest the latest Pi build embedded in Cyberful
make typecheck  # run policy and TypeScript checks
make test       # run the complete test suite
make build      # build standalone release binaries
make run        # launch Cyberful from source

make build resolves the current official ATT&CK releases once, validates and indexes the three domains, and embeds the resulting snapshot in every standalone binary. Source launches such as make run never download ATT&CK; they use the snapshot left by the latest successful build and report DATASET_UNAVAILABLE when none is available.

Source development requires Bun 1.3.14, Node.js 24 with npm, Python 3.10+, and Docker. See the contributing guide for the complete build, test, runtime, and release workflow.

Configuration and local state

Cyberful creates a secret-free settings.yaml in the launch directory. It defines providers, models, fallback routing, delegation, trusted instructions, and the progressive skill-catalog budget; provider credentials stay in Cyberful's protected credential store or named environment variables. See Agent providers and fallback and .env-example.

Workareas live under work/<name>/ and session logs under logs/session-logs/. They can contain sensitive evidence. Never commit workareas, transcripts, browser profiles, ZAP or Ghidra state, generated reports, credentials, or tokens.

The first local engagement also prepares the release-pinned CVE Dictionary in the foreground. Before downloading, Cyberful checks an explicit verified path, the managed pointer, verified orphan snapshots, and a source checkout's fixed dist/cve-dictionary directory; selecting an orphan repairs the missing pointer atomically. Release 2026.08.05 downloads about 5.18 GiB only when no verified local candidate exists, expands to about 24.47 GiB, requires approximately 31 GiB of additional free space during installation, and shows verified download and activation progress on stderr. Later startups reuse the local snapshot without a network update check.

Start a specific headless workflow with cyberful run --workflow bug-bounty --workarea <name> "<objective>". Add --format json for the redacted versioned coding-agent stream, inspect durable or attached state with cyberful session inspect, and supervise an attached TUI with cyberful session watch. The public cross-agent skill and full control-room contract are documented in Cyberful Skill: How to integrate cyberful in your coding agent. Resume an existing session with cyberful run --continue or cyberful run --session <id>; session lifecycle, steering, approvals, reports, and cleanup are covered in Sessions, configuration, and reports.

Documentation

• Cyberful documentation
• Your first penetration test
• Application security workflows
• Terminal interface
• Cyberful Skill: How to integrate cyberful in your coding agent
• Architecture
• Execution model
• Security runtimes
• Built-in skill catalog
• CVE Dictionary
• CVE Dictionary technical README
• MITRE ATT&CK MCP
• Contributing
• Security policy

Support Cyberful

Love Cyberful? Give us a ⭐ on GitHub!

Acknowledgements

Cyberful builds on the incredible work of open-source projects like Pi, OpenTUI, the Cyberful agent-browser fork and its upstream project, Chromium, OWASP ZAP, Ghidra, and the MITRE ATT&CK knowledge base. Huge thanks to their maintainers!

License and responsible use

Cyberful is released under the GNU Affero General Public License v3.0 only (AGPL-3.0-only). Use it only on systems and source you are authorized to assess. You are responsible for scope, authorization, legal compliance, and safe operation.