The Problem

Security teams juggle dozens of point tools—Nmap for recon, Metasploit for exploitation, ZAP for web testing—with no unified way to orchestrate them, preserve evidence, or verify that a finding is actually exploitable. Results live in scattered notes, and remediation tracking is manual. Cyberful addresses this by wrapping 316 tools in a multi-agent system with a deterministic control plane that enforces scope and produces verified, reproducible reports.

What This Does

Cyberful is an AI-driven red teaming platform. The core engine lives in cyberful/src/, with a CLI entry point at cyberful/src/cli/cmd/agent.ts. The system uses "hypothesis workers" to map attack surfaces and test exploit paths, then a control plane validates findings independently before they reach a report.

The repo ships with a substantial skill library under cyberful/builtin/skills/—each skill (e.g., analyze-http-traffic-evidence, assess-application-threat-model) bundles a SKILL.md spec, agent configs, JSON schemas, Python scripts, and tests. There are also prebuilt agent workflows for pentesting, code audits, and bug bounty in cyberful/builtin/agents/. The mcps/ directory contains Model Context Protocol servers, including a Dockerized cyberful-os runtime and a ZAP integration.

How It Is Wired

Execution starts at cyberful/src/cli/cmd/agent.ts, which routes through the command handlers in cyberful/src/cli/cmd/. The system uses a message bus (cyberful/src/bus/index.ts) to coordinate agent activity. The control plane is deterministic—it doesn't rely on LLM judgment for scope enforcement or evidence preservation. The runtime is containerized via mcps/cyberful-os/Dockerfile and docker-compose.yml, which bundle ZAP, Ghidra, Python, Ruby, and the offensive toolchain.

The cyberful/builtin/agents/ directory defines workflows: pentest/, code-audit/, and bug-bounty/ each have recon.md, exploit.md, verify.md, and report.md phases. These route through the agent CLI, which spawns skills and tools. The budgets.json files in each agent directory cap resource usage per phase.

The wiring for the full control flow—how a skill's Python script calls out to external tools, how evidence schemas are validated, how reports are generated—is not fully mapped in this analysis. The ARCHITECTURE.md and CODE.md files in the root likely document this, but the internal call graph wasn't traced.

How To Use It

npm install --global cyberful
mkdir cyberful-lab && cd cyberful-lab
# Create settings.yaml based on example.settings.yaml
cyberful auth login
cyberful

Requirements per the README: Node.js 18+, npm, Docker with Linux containers, a configured model provider, 100 GB disk, 10 GB RAM for Docker. The Makefile and package.json at root suggest additional build targets, but the README's npm flow is the documented path.

Real-World Use

For an authorized pentest engagement: create a lab folder, configure settings.yaml with your scope and model provider, run cyberful, select the Pentest workflow. The system runs recon (Nmap, Nuclei), exploitation (Metasploit), and verification—each finding must be independently reproduced before it's reported. The ZAP proxy integration in mcps/zap/ handles web traffic, and the cyberful-os Docker runtime provides the isolated environment.

Code Health & Issues

Static analysis found no structural red flags: 487 test files, CI via .github/workflows/ci.yml, a LICENSE (AGPL-3.0), and lockfiles (bun.lock, package-lock.json) are all present. The repo is heavily TypeScript (635 files) with 100 Python files for skill scripts.

  • Low - Repo is a fork of cyberful/cyberful (84 stars) but this clone has 0 stars; verify upstream maintenance status before committing.
  • Low - The bun.lock and bunfig.toml at root suggest Bun usage, but the README documents npm only; the build path may be ambiguous.

The Bottom Line

Cyberful is a serious, well-structured attempt at AI-driven security testing with real tool integration and evidence preservation. The Docker requirement and 100 GB disk minimum make it heavy for casual use, but for teams doing authorized pentests or bug bounty work, the deterministic control plane and verification pipeline are genuinely useful. It's not a toy—but verify the upstream project's activity before building on it.