27 scope-enforced AI agents that run the full pentest kill-chain (recon → exploit → post-ex → DFIR → report) as a one-command Claude Code plugin. Backed by 754 MITRE-mapped skills.

For authorized security testing only. Every network command is scope-gated by scope_check.py — violations are blocked deterministically at the OS level, not by convention or prompt instruction.

Most AI pentest tools stop at exploit. ThreatSwarm runs the complete kill chain — recon, exploitation, post-exploitation, DFIR, and a CVSS-scored report — in a single session. It ships as a one-command Claude Code plugin (no Docker, no Postgres, no cloud account required), enforces scope on every tool invocation across all 27 agents, and loads its methodology from the 754-skill Anthropic-Cybersecurity-Skills library mapped to ATT&CK, CSF 2.0, ATLAS, D3FEND, and AI RMF.

$ claude --plugin-dir ./threatswarm-plugin

> /threatswarm:engage 10.10.10.5

  ✓  scope check  —  10.10.10.5 found in scope.txt

  →  recon agent starting...

[*] nmap -sS -T3 -p- 10.10.10.5
    22/tcp   open  ssh      OpenSSH 8.9p1
    80/tcp   open  http     Apache httpd 2.4.51
    8080/tcp open  http     Apache Tomcat 9.0.45
    3306/tcp open  mysql    MySQL 8.0.28

[*] nuclei -u http://10.10.10.5:8080 -severity critical,high
    [critical]  CVE-2021-44228  Log4Shell RCE         http://10.10.10.5:8080
    [high]      CVE-2020-1938   Tomcat AJP Ghostcat   http://10.10.10.5:8080

[*] feroxbuster -u http://10.10.10.5 -x php,txt
    200  /admin          Apache Tomcat Manager  [auth required]
    200  /api/v1/users   [JSON, 247 entries]
    200  /backup.zip     [3.2 MB]

  Attack vectors — prioritised:
  1  [CRITICAL]  Log4Shell RCE  →  /threatswarm:attack 10.10.10.5 exploit CVE-2021-44228
  2  [HIGH]      API user enum  →  /threatswarm:attack 10.10.10.5 api
  3  [HIGH]      Backup file    →  /threatswarm:attack 10.10.10.5 web

> /threatswarm:attack 10.10.10.5 exploit CVE-2021-44228

  →  exploit agent  [model: claude-opus-4-7]
  →  vuln-researcher confirming PoC reliability... weaponized ✓
  →  listener started on 10.10.14.1:4444
  →  payload delivered via X-Api-Version header
  [+] shell obtained  —  www-data@10.10.10.5

> /threatswarm:pwned 10.10.10.5

  →  post-ex agent
  [*] sudo -l  →  (ALL : ALL) NOPASSWD: /usr/bin/python3
  [+] privesc  →  root@10.10.10.5
  [*] /etc/shadow extracted  →  evidence/20260429/10.10.10.5/creds/shadow.hashes
  [*] crackmapexec smb 10.10.10.0/24  →  3 hosts reachable with root hash

> /threatswarm:report 10.10.10.5
  →  report-writer agent  →  reports/10.10.10.5_pentest.md  ✓

Scope enforced throughout. Evidence written to evidence/20260429/10.10.10.5/.

Marketplace (recommended — installs into any project):

/plugin marketplace add mukul975/ThreatSwarm
/plugin install threatswarm@threatswarm

Standalone — clone and use directly:

git clone https://github.com/mukul975/ThreatSwarm && cd ThreatSwarm
claude

Single session — load without installing:

git clone https://github.com/mukul975/ThreatSwarm
claude --plugin-dir ./ThreatSwarm/threatswarm-plugin

1. Add authorized targets to scope.txt:

   10.10.10.5
   10.10.10.0/24
   example.com
   

2. Set your attacker IP in .claude/settings.json → "LHOST": "10.10.14.1"
3. Run claude and start a kill chain:

   /threatswarm:engage 10.10.10.5    # recon → ranked attack vectors
   /threatswarm:attack 10.10.10.5 web  # route to specialist agent
   /threatswarm:pwned 10.10.10.5     # post-shell: privesc → creds → lateral
   /threatswarm:hunt "C2 beaconing"  # ATT&CK-mapped threat hunt
   /threatswarm:ir ransomware        # DFIR triage
   /threatswarm:report engagement    # CVSS-scored PDF report
   

scope_check.py runs as a PreToolUse hook before every Bash command. It extracts IPs and hostnames from the command, resolves them against scope.txt with full CIDR awareness, and returns exit code 2 to block out-of-scope commands. Claude cannot override this — the hook executes outside the agent loop.

scope.txt example:

# Authorized targets — add before starting any engagement
10.10.10.5
10.10.10.0/24
*.example.com
api.acme-staging.io

scope.yaml example (HackerOne RoE compatible):

# ThreatSwarm scope file — maps directly from HackerOne program rules
targets:
  in_scope:
    - type: cidr
      value: "10.10.10.0/24"
      note: "Internal lab network"
    - type: domain
      value: "*.example.com"
      note: "All subdomains in scope"
    - type: ip
      value: "192.168.1.100"
  out_of_scope:
    - "prod.example.com"
    - "10.10.10.254"   # firewall — do not touch
engagement:
  type: "blackbox"
  authorized_by: "Jane Smith, CISO"
  authorization_date: "2026-04-29"
  rules_of_engagement: "No DoS, no data exfiltration, stop-and-report on critical finds"

Three enforcement layers run in parallel.

Three-layer safety model:

Layer 1 — scope_check.py (deterministic hook) Runs before every Bash command. Extracts IPs and hostnames, checks against scope.txt with CIDR awareness. Exit code 2 = command blocked. Executes outside the agent loop — Claude cannot override it.

Layer 2 — CLAUDE.md rules Never run active tools in the main thread. Always delegate to agents. Never store plaintext credentials. Never exfiltrate PII. Enforced in every session.

Layer 3 — Path-scoped output rules evidence/** — mandatory ATT&CK TTP fields, no raw credentials
reports/** — CVSS 3.1 vector required per finding, credentials redacted to [REDACTED]
loot/** — hash + location reference only

Every agent loads its methodology from Anthropic-Cybersecurity-Skills (5,700+ ⭐) — 754 structured skills mapped to MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, and AI RMF. The skills library is the reason ThreatSwarm agents follow professional methodology rather than ad-hoc prompting.

evidence/
└── 20260429/
    └── 10.10.10.5/
        ├── nmap/           raw scanner output
        ├── nuclei/         vulnerability findings
        ├── web/            HTTP responses, screenshots
        ├── creds/          hashes only — no plaintext ever
        └── screenshots/
findings.md                 CVSS-scored findings, auto-aggregated

findings_sync.py runs at session end and aggregates CRITICAL/HIGH findings to evidence/FINDINGS_SUMMARY.md.

./scripts/worktree_setup.sh acme-external acme-internal
# Each worktree gets its own scope.txt, evidence/, and git branch
cd ../ThreatSwarm-acme-external && claude

• Written authorization required for every target
• Social engineering campaigns need signed RoE
• OT/ICS: passive monitoring only unless explicitly scoped
• No DoS-class attacks without explicit written authorization
• See CONTRIBUTING.md for responsible disclosure policy

@software{jangra2026threatswarm,
  author    = {Jangra, Mahipal},
  title     = {{ThreatSwarm}: Scope-Enforced Multi-Agent Penetration Testing
               with the {Claude Code} Plugin Substrate},
  year      = {2026},
  url       = {https://github.com/mukul975/ThreatSwarm},
  note      = {27-agent full kill-chain pentesting plugin for Claude Code}
}

arXiv preprint forthcoming.

• scope.yaml HackerOne RoE auto-converter
• CI/CD integration (GitHub Actions, GitLab CI, SARIF output)
• Burp Suite / Caido extension
• GenAI attack surface agent
• Supply chain / dependency confusion agent
• IPv6 support in scope_check.py
• Community agent bounty program (Nuclei-template model)

See CONTRIBUTING.md. Priority areas: new agents for emerging surfaces (GenAI attacks, supply chain, LLM injection), IPv6 scope support, CI pipeline examples. Open an issue before starting major work.

• Claude Code 1.0.33+
• Python 3.8+ (hook scripts)
• Kali Linux or equivalent (for pentest tools)
• Anthropic-Cybersecurity-Skills recommended

MIT License · Built by Mahipal Jangra (@mukul975)

Powered by Anthropic-Cybersecurity-Skills · Privacy-Data-Protection-Skills