PentAGI

PentAGI is a fully autonomous AI red-team tool with 8,200+ GitHub stars that has drawn significant attention across the cybersecurity community. Open source and actively maintained, it automates the complete offensive security lifecycle — from passive reconnaissance through exploitation and post-exploitation — with zero human intervention required at any stage.

Installation

Docker (Recommended)

# Pull and run latest release
docker pull pentagi/pentagi:latest

# Run with API key and results volume
docker run -it --rm \
  -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY \
  -v $(pwd)/results:/app/results \
  -p 8080:8080 \
  pentagi/pentagi:latest

# Run with web dashboard
docker run -d \
  -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY \
  -v $(pwd)/results:/app/results \
  -v $(pwd)/config:/app/config \
  -p 8080:8080 \
  --name pentagi \
  pentagi/pentagi:latest

# Open dashboard: http://localhost:8080

Docker Compose

git clone https://github.com/pentagi/pentagi
cd pentagi

# Copy and configure environment
cp .env.example .env
# Edit .env with your API keys

# Start full stack (agent + dashboard + DB)
docker-compose up -d

# Check logs
docker-compose logs -f pentagi

From Source

git clone https://github.com/pentagi/pentagi
cd pentagi

# Install dependencies
pip install -r requirements.txt

# Install optional security tool dependencies
./scripts/install-tools.sh

# Run
python -m pentagi

pip

pip install pentagi

# Verify
pentagi --version

Configuration

Environment Variables (`.env`)

# Required: AI provider
ANTHROPIC_API_KEY=sk-ant-...
PENTAGI_MODEL=claude-opus-4-5        # Primary model
PENTAGI_FAST_MODEL=claude-haiku-4-5  # Used for sub-tasks

# Optional: additional providers
OPENAI_API_KEY=sk-...
OPENAI_MODEL=gpt-4o

# Intelligence sources
SHODAN_API_KEY=...
CENSYS_API_KEY=...
HUNTER_API_KEY=...                    # Email OSINT

# Web dashboard
PENTAGI_WEB_HOST=0.0.0.0
PENTAGI_WEB_PORT=8080
PENTAGI_WEB_AUTH_TOKEN=your-secret-token

# Output
PENTAGI_RESULTS_DIR=/app/results
PENTAGI_LOG_LEVEL=info

Operation Configuration (`config/operation.yaml`)

operation:
  name: "target-assessment"
  description: "Full autonomous red team assessment"

target:
  primary: "example.com"
  ip_ranges:
    - "10.0.0.0/8"
  exclude_hosts:
    - "10.0.0.1"
  exclude_ports:
    - 22                    # Skip SSH brute-force

phases:
  recon: true
  vulnerability_scan: true
  exploitation: true
  post_exploitation: true
  reporting: true

constraints:
  max_duration: 86400       # 24 hours max
  max_parallel_tasks: 5
  safe_mode: true           # Confirm before destructive actions
  stealth_mode: false       # Slow + quiet vs fast + noisy

output:
  format: html              # html | json | markdown | pdf
  include_evidence: true
  include_timeline: true
  include_remediation: true

Core Commands

Command	Description
`pentagi start`	Start autonomous operation with configured target
`pentagi start --target <host>`	Start operation against specific target
`pentagi start --config <file>`	Start with a configuration file
`pentagi status`	Show current operation status
`pentagi status --op <id>`	Show status of a specific operation
`pentagi pause <op-id>`	Pause a running operation
`pentagi resume <op-id>`	Resume a paused operation
`pentagi abort <op-id>`	Safely abort an operation
`pentagi recon <target>`	Run recon phase only
`pentagi scan <target>`	Run vulnerability scan phase only
`pentagi exploit <target>`	Run exploitation phase only
`pentagi post-exploit <session>`	Run post-exploitation phase
`pentagi report <op-id>`	Generate final operation report
`pentagi history`	List all previous operations
`pentagi dashboard`	Open web dashboard in browser
`pentagi update`	Update tool integrations and AI prompts

Advanced Usage

Autonomous Operation Lifecycle

# Start a full autonomous assessment
pentagi start \
  --target example.com \
  --name "example-assessment-2025" \
  --phases all \
  --max-duration 24h \
  --output ./results/

# Monitor in real time via dashboard
pentagi dashboard

# Or monitor via CLI
watch -n 10 pentagi status --op example-assessment-2025

# Tail operation logs
pentagi logs --op example-assessment-2025 --follow

Phase-by-Phase Execution

Phase	Key Activities
Recon	OSINT, subdomain enum, port/service scan, tech fingerprinting
Vulnerability Scan	CVE correlation, nuclei templates, misconfig detection
Exploitation	Automated exploit selection, payload delivery, shell capture
Post-Exploitation	Privilege escalation, lateral movement, data collection
Reporting	Evidence compilation, risk rating, remediation guidance

# Run phases selectively
pentagi start \
  --target example.com \
  --phases "recon,vulnerability_scan" \
  --skip-phases "exploitation,post_exploitation" \
  --name "non-intrusive-scan"

# Resume from a specific phase
pentagi start \
  --resume-from exploitation \
  --op "example-assessment-2025"

Stealth Mode Operations

# Low-and-slow operation for blue team evasion
pentagi start \
  --target example.com \
  --stealth-mode \
  --timing "sneaky" \       # paranoid | sneaky | polite | normal | aggressive
  --randomize-user-agents \
  --rotate-source-ips \
  --c2-jitter 60 \
  --name "stealth-op-2025"

Custom Tool Integration

Built-in Tools	Category
nmap, masscan, zmap	Port scanning
amass, subfinder, assetfinder	Subdomain enumeration
nuclei, nikto, wapiti	Vulnerability scanning
sqlmap, commix	Injection exploitation
hydra, medusa	Credential attacks
metasploit, exploitdb	Exploitation frameworks
linpeas, winpeas, pspy	Post-exploitation enumeration
mimikatz, secretsdump	Credential harvesting

# Add a custom tool to the PentAGI toolset
pentagi tools register \
  --name "custom-scanner" \
  --binary /usr/local/bin/custom-scanner \
  --description "Proprietary vulnerability scanner" \
  --input-schema ./schemas/custom-scanner-input.json

# List all registered tools
pentagi tools list

# Test a tool integration
pentagi tools test custom-scanner --target example.com

Web Dashboard Features

# Start with dashboard enabled
pentagi start --target example.com --web-ui

# Dashboard provides:
# - Real-time operation timeline
# - Live agent decision log
# - Interactive attack graph visualization
# - Found credentials and loot browser
# - One-click report generation
# - Operation comparison across engagements

# Access at: http://localhost:8080
# API at:    http://localhost:8080/api/v1

API Usage

# Start operation via API
curl -X POST http://localhost:8080/api/v1/operations \
  -H "Authorization: Bearer $PENTAGI_WEB_AUTH_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "api-initiated-op",
    "target": "example.com",
    "phases": ["recon", "vulnerability_scan"],
    "safe_mode": true
  }'

# Get operation status via API
curl http://localhost:8080/api/v1/operations/api-initiated-op \
  -H "Authorization: Bearer $PENTAGI_WEB_AUTH_TOKEN"

# Stream live logs via WebSocket
wscat -c "ws://localhost:8080/api/v1/operations/api-initiated-op/logs" \
  -H "Authorization: Bearer $PENTAGI_WEB_AUTH_TOKEN"

Common Workflows

Full Autonomous Red Team Assessment

# Step 1: Configure target scope
cat > config/my-engagement.yaml << 'EOF'
target:
  primary: "example.com"
  ip_ranges: ["10.0.0.0/24"]
phases:
  recon: true
  vulnerability_scan: true
  exploitation: true
  post_exploitation: true
  reporting: true
constraints:
  safe_mode: true
  max_duration: 86400
output:
  format: html
  include_remediation: true
EOF

# Step 2: Launch operation
pentagi start \
  --config config/my-engagement.yaml \
  --name "full-assessment-$(date +%Y%m%d)"

# Step 3: Monitor (optional — it's fully autonomous)
pentagi status --op "full-assessment-$(date +%Y%m%d)" --watch

# Step 4: Collect report when done
pentagi report \
  --op "full-assessment-$(date +%Y%m%d)" \
  --format html \
  --output ./reports/full-assessment.html

Bug Bounty Hunting

# Configure for bug bounty scope
pentagi start \
  --target "*.example.com" \
  --scope-file ./scope.txt \
  --phases "recon,vulnerability_scan" \
  --focus "xss,sqli,ssrf,idor,open-redirect" \
  --severity-threshold medium \
  --output ./bb-results/ \
  --name "bb-example-$(date +%Y%m%d)"

# Export findings as bug bounty report format
pentagi report \
  --op "bb-example-$(date +%Y%m%d)" \
  --format markdown \
  --template bug-bounty \
  --output submission-draft.md

Continuous Automated Assessment

# Run weekly automated assessment
cat > weekly-assess.sh << 'EOF'
#!/bin/bash
OP_NAME="weekly-$(date +%Y-%W)"
pentagi start \
  --config /etc/pentagi/prod-scope.yaml \
  --name "$OP_NAME" \
  --phases "recon,vulnerability_scan" \
  --diff-baseline ./baselines/last-week.json \
  --alert-new-findings \
  --alert-email security@example.com

pentagi report \
  --op "$OP_NAME" \
  --format html \
  --output /reports/$OP_NAME.html
EOF
chmod +x weekly-assess.sh
# cron: 0 1 * * 1 /path/to/weekly-assess.sh

Competitive CTF Automation

# Use PentAGI for CTF challenge automation
pentagi start \
  --target "ctf-machine.hackthebox.eu" \
  --mode ctf \
  --objective "get root flag" \
  --time-limit 4h \
  --aggressive \
  --name "htb-machine-name"

Tips and Best Practices

Use safe-mode for production targets — Always set safe_mode: true for any target with production traffic; this gates destructive exploits behind a human confirmation prompt while still running all recon and vulnerability discovery autonomously.

Resource-size your Docker container — PentAGI’s parallel task execution is CPU/memory-intensive; allocate at minimum 4 CPU cores and 8GB RAM (--cpus 4 --memory 8g) for smooth multi-phase operation.

Baseline your assessments — Use --diff-baseline to compare against a previous operation’s findings, so recurring issues are separated from newly discovered vulnerabilities in reports.

Scope files prevent scope creep — Always provide a --scope-file for bug bounty engagements; the AI agent will otherwise follow every link and subdomain it finds, potentially going out of scope.

Dashboard over CLI for complex ops — The web UI provides an interactive attack graph that makes lateral movement paths and exploit chains far easier to understand than raw log output.

CTF mode vs engagement mode — Use --mode ctf for Hack The Box/TryHackMe machines (aggressive, no scope restrictions); use the default engagement mode with safe_mode: true for real-world client assessments.

Export evidence during the op — Don’t wait until the operation completes to collect screenshots and tool output; use pentagi evidence export --op <id> mid-operation to save artifacts if connectivity is unstable.

Watch agent decisions — Enable LOG_LEVEL=debug on first use to understand how the AI reasons about target selection and exploit prioritization — invaluable for calibrating the tool to your engagement style.