Skip to content

Testing with Attackers

Jump to bottom
Stephen edited this page Mar 5, 2026 · 4 revisions

Testing with Attackers

LABYRINTH is designed to capture, degrade, and commandeer autonomous offensive AI agents. To test it, you need an attacker agent pointed at the portal trap services.

All attacker agents run inside Docker containers, isolated from your host machine and connected to the LABYRINTH network.

Quick Start

labyrinth deploy -t                     # Deploy the portal trap
labyrinth bait drop                     # Plant randomized bait credentials
labyrinth attacker setup                # Pick and configure an attacker agent

The deploy command will prompt you to drop bait automatically. If you skip it, run labyrinth bait drop before launching an attacker — agents need discoverable credentials to enter the trap.

Full Lifecycle

labyrinth deploy -t                     # 1. Deploy the portal trap
labyrinth bait drop                     # 2. Plant randomized bait credentials
labyrinth attacker setup                # 3. Pick an agent and configure
labyrinth attacker run pentagi          # 4. Quick-launch (or re-launch) an agent
labyrinth tui                           # 5. Watch captures in real time
labyrinth attacker stop pentagi         # 6. Stop when done
labyrinth bait clean                    # 7. Remove bait credentials
labyrinth attacker uninstall pentagi    # 8. Clean up agent

The Bait Trail

Agents need breadcrumbs to find their way into the trap. labyrinth bait drop generates a unique randomized identity (company, users, passwords, API keys) and plants discoverable files on the portal trap services:

Agent scans HTTP → discovers /robots.txt → finds sensitive paths
  → reads /.env, /backup/credentials.csv → extracts SSH credentials
    → logs into SSH with discovered creds → enters the trap (L1)
      → finds bait files inside SSH → triggers escalation (L2)
        → contradictions erode the agent's world model (L2)
          → terminal corruption blinds the agent (L3)
            → API interception captures/rewrites instructions (L4)

Each bait drop generates a unique identity so the portal trap cannot be fingerprinted — no two deployments share the same credentials.

labyrinth bait show                     # View current bait credentials
labyrinth bait clean                    # Remove all bait when done

Agent Status

labyrinth attacker list                 # Quick overview table
labyrinth attacker status               # Detailed view of installed agents

Attacker Agents

1. PentAGI — Fully Autonomous Multi-Agent

Source: github.com/vxcontrol/pentagi

Best for: Hands-off autonomous pentesting. Deploy it and watch.

Feature Detail
Interface Web UI at https://localhost:8443
Tools 20+ built-in (nmap, metasploit, sqlmap, nikto, etc.)
Isolation Full Docker sandboxing
LLM OpenAI, Anthropic, Gemini, Bedrock, or Ollama
Deployment Docker Compose stack in ~/.labyrinth/attackers/pentagi/

PentAGI runs a multi-agent system where specialized AI roles (researcher, developer, executor) coordinate autonomously. It has its own Docker Compose stack managed by the CLI.

Setup:

labyrinth attacker setup                # Select PentAGI
# or
labyrinth attacker run pentagi          # Quick-launch (will prompt for setup if needed)

Prompt examples (in PentAGI web UI):

Penetration test the SSH service at labyrinth-ssh:22
Penetration test the web app at http://labyrinth-http:80

Management:

labyrinth attacker stop pentagi         # Stop the compose stack
labyrinth attacker run pentagi          # Restart it
labyrinth attacker uninstall pentagi    # Remove everything (containers, images, config)

2. PentestAgent — Interactive AI Pentesting with TUI

Source: github.com/GH05TCREW/PentestAgent

Best for: Guided pentesting with interactive control and playbooks.

Feature Detail
Interface Terminal TUI
Modes Agent (autonomous), Crew (multi-agent), Assist (chat)
Tools nmap, netcat, curl (base); full Kali suite in Kali image
LLM Any via LiteLLM (OpenAI, Anthropic, Google, Ollama)
Container labyrinth-attacker-pentestagent

Setup:

labyrinth attacker setup                # Select PentestAgent

Commands inside the TUI:

/agent Pentest SSH at labyrinth-ssh:22
/agent Pentest web app at http://labyrinth-http:80
/crew Full pentest of labyrinth-ssh:22 and http://labyrinth-http:80
/target labyrinth-ssh
/tools
/quit

3. Strix — AI Hacker Agents

Source: github.com/UseStrix/strix

Best for: Web application security testing.

Feature Detail
Interface CLI with TUI
Sandbox Kali Docker container (auto-managed)
Focus Web app vulnerabilities
LLM Any via LiteLLM

Strix runs as a host-installed CLI that launches its own Docker sandbox containers. The labyrinth attacker setup command pulls the sandbox image and prints host-side installation instructions.

Setup:

labyrinth attacker setup                # Select Strix — pulls sandbox, shows install steps

Usage (after installing Strix on host):

export STRIX_LLM=openai/gpt-4o
export LLM_API_KEY=<your-key>

strix --target http://localhost:8080
strix --target localhost --instruction "Pentest SSH on port 22"

4. Custom Agent / Manual Testing

Best for: Running your own tools, manual testing, or agents not listed above.

The CLI launches a Kali Linux container directly on the LABYRINTH network with common tools pre-installed (nmap, hydra, curl, wget, netcat, sqlmap, nikto, dirb, sshpass).

Setup:

labyrinth attacker setup                # Select Custom Kali
# or
labyrinth attacker run kali             # Quick-launch directly

Inside the container:

nmap -sV labyrinth-ssh
ssh root@labyrinth-ssh
curl http://labyrinth-http
hydra -l root -P /usr/share/wordlists/rockyou.txt ssh://labyrinth-ssh
nikto -h http://labyrinth-http

Bring Your Own Agent

Connect any Docker container to the LABYRINTH network:

# Run a new container on the network
docker run -it --rm \
    --network labyrinth-net \
    your-image:tag

# Or connect an existing container
docker network connect labyrinth-net your-container

Target Addresses (from inside the network)

Service Address
SSH Portal Trap labyrinth-ssh:22
HTTP Portal Trap labyrinth-http:80
Dashboard labyrinth-dashboard:9000
Orchestrator labyrinth-orchestrator

Monitoring During Tests

While an attacker agent is running, monitor captures in real time:

# Terminal dashboard
labyrinth tui

# Web dashboard
open http://localhost:9000

Generate a Post-Test Report

After the attacker finishes, generate a forensic report:

labyrinth report                           # Latest session (terminal)
labyrinth report --format md -o report.md  # Markdown with attack graph
labyrinth report --format json             # JSON export

Reports include MITRE ATT&CK timeline mapping, credential analysis, service interaction logs, Mermaid attack graphs, and effectiveness assessment. See Forensics-and-API for details.

Teardown

# Stop a specific agent
labyrinth attacker stop pentagi

# Stop all agents
labyrinth attacker stop --all

# Full uninstall (remove containers, images, config)
labyrinth attacker uninstall --all

Safety Notes

  • All attacker agents run inside Docker containers, not on your host
  • Agents are connected to the labyrinth-net bridge network
  • The LABYRINTH portal trap services are isolated in their own containers
  • No attacker traffic touches your host network or other services
  • API keys are never stored on disk — read from environment variables or prompted each time
  • Agent configuration is stored in ~/.labyrinth/attackers/ (user-global)
  • Always use labyrinth attacker stop or labyrinth attacker uninstall to clean up after testing

Attacker Agent Links

Agent Repository Documentation
PentAGI github.com/vxcontrol/pentagi PentAGI Docs
PentestAgent github.com/GH05TCREW/PentestAgent PentestAgent Docs
Strix github.com/UseStrix/strix Strix Docs
Kali Linux kali.org Kali Docker

Related Pages

Project LABYRINTH

Home

Getting Started

Architecture

Layers

Operations

Clone this wiki locally