Pentest Swarm AI is an autonomous, open-source API & web-app pentester built on a real swarm. Dozens of AI agents coordinate through a shared stigmergic blackboard to exploit, chain, and prove vulnerabilities at machine speed.
- 🐝 A real swarm, not a pipeline — dozens of agents work your surface concurrently, coordinating via pheromone-weighted findings; attack chains emerge instead of being scripted.
- 💥 Exploits, doesn't just scan — chains BOLA/IDOR, JWT forgery, mass assignment, SSRF, SQLi/NoSQLi, GraphQL abuse — then captures evidence and writes the report.
- 🧠 Any model, self-hosted — Claude, Together AI (GLM/Qwen/DeepSeek), Gemini, OpenAI-compatible, or fully local Ollama / LM Studio. We're the harness that gives models hands.
- ✅ Proof over maybes — non-LLM verification oracles (OOB / differential / timing) confirm findings against the live target; proven ones get a VERIFIED badge.
- 🛡️ Safe by construction — fail-closed scope, egress guard, an authorization gate, and a full audit trail.
- 🎯 Drop-in use cases — bug bounty, external ASM, CI/CD, internal network, CTF, GraphQL, OWASP Top 10 — each a one-command playbook.
Hack at machine speed.
📖 New here? Start with the documentation — quickstart, provider setup, CLI reference, and troubleshooting. (Or run pentestswarm docs to open it.)
XBOW proved a point: AI can top the bug-bounty leaderboards. But XBOW is a closed, hosted SaaS — one agent, on their cloud, on their model, at their price. We think the future of offense is open, self-hosted, and a swarm — so we built it.
Pentest Swarm vs. XBOW
| Pentest Swarm AI | XBOW |
|---|
| Open source | ✅ AGPL, fork it | ❌ closed |
| Self-hosted | ✅ your infra, your model, $0 floor | ❌ their cloud, their bill |
| A real swarm | ✅ dozens of agents, concurrent | ❌ single agent |
Three structural wins — and the swarm is the one that compounds. This isn't a single planner LLM walking recon → classify → exploit → report down a fixed line. Dozens of agents work your surface at once, coordinating through a shared stigmergic blackboard: the instant a finding lands it wakes whichever agent it's relevant to, so a 1,000-subdomain target gets hit in parallel — attacking at machine speed. And unlike a scanner that spits out unverified "maybes," the swarm exploits what it finds, proves it with captured evidence, then writes the report.
Who it's for: pentesters covering a whole scope overnight · bug bounty hunters racing to first-blood on a fresh target · red teamers who need breadth fast · researchers pushing autonomous offense.
Run it on any model — Claude, anything OpenAI-compatible (incl. Together AI's hosted Llama/Qwen/DeepSeek), the new security-tuned open models (Pentest-R1 and the wave behind it), or fully local Ollama / LM Studio. We don't compete with those models — we're the harness that gives them hands: real tools, swarm coordination, scope safety, and evidence-backed reports. Air-gapped, zero API cost, and not one byte of your data leaving your box.
For authorized testing only — see the full legal disclaimer and Responsible Use Policy.
What makes this a swarm?
Most "multi-agent" pentesting tools are a single planner LLM dispatching to specialist agents in a fixed order — recon → classify → exploit → report. That's a pipeline, not a swarm.
Pentest Swarm AI is built around three swarm-intelligence primitives:
- Stigmergy — agents coordinate by reading and writing findings on a shared blackboard, not by a central planner telling them what to do. A finding's pheromone weight biases other agents toward it and decays over time, so stale paths die naturally.
- Emergence — attack chains appear that no single agent planned. A recon finding wakes the classifier; a high-severity classification wakes the exploit agent; exploit results feed back into the board and wake the report agent. Order isn't prescribed — it emerges from the blackboard state.
- Decentralization — each agent runs its own trigger predicate. Add a new agent with its own predicate and it joins the swarm without anyone rewriting the orchestrator.
Why it wins: a pipeline is sequential (one stage at a time), can't fold a mid-run discovery back into recon, and is bounded by its slowest stage. A swarm works your surface concurrently, reacts to every new finding, lets attack chains emerge instead of scripting them, and lets stale paths decay and die. The payoff on stage is breadth — a long list of proven findings, because many agents chew on the surface at once.
📖 Full write-up: Architecture — Swarm vs. Pipeline (the blackboard, pheromone decay, emergent chains, and why concurrency stays cheap). Diagrams also live in docs/.
We built this because the category was empty. Every tool marketed as "swarm" was actually a pipeline. If you find a counter-example, open an issue — we'll add them to the comparison table.
Quick Start
One command to install. One command to run. No API key, no cloud, no bill to start.
1 · Install (pick one)
npm install -g @armurai/pentestswarm # npm (Node ≥16)
curl -fsSL https://raw.githubusercontent.com/Armur-Ai/Pentest-Swarm-AI/main/scripts/install.sh | sh # any macOS/Linux
brew install Armur-Ai/tap/pentestswarm # Homebrew
go install github.com/Armur-Ai/Pentest-Swarm-AI/cmd/pentestswarm@latest # Go toolchain
docker run --rm ghcr.io/armur-ai/pentestswarm:latest --help # Docker
2 · Run it
pentestswarm run
That's it. run opens the interactive TUI launcher — no flags to memorize:
- Pick your AI provider and paste a key right in the UI — Together AI (hosted Llama / Qwen / DeepSeek), Claude, OpenAI, Gemini, or fully local Ollama / LM Studio (no key at all).
- Point it at a target — a URL, or a bundled vulnerable lab (crAPI, Juice Shop, VAmPI, DVGA) that spins up, gets attacked, and tears down after. Legal, safe, zero setup.
- Choose your live view — a web dashboard on
localhost:7777 and a full-screen terminal TUI with live charts, a swarm-topology diagram, and a graded findings stream. Both, by default.
- Then launch and watch the swarm work at machine speed.
A readiness check runs right in the launcher (Go, Docker, tools, provider) — it never blocks; anything missing is shown as a note you can fix or ignore.
Prefer flags? (scripting / CI)
run just wraps scan, so everything is scriptable too:
# watch it find a real vuln in ~2 min — bundled lab, local model, no key
pentestswarm scan --lab --lab-target crapi --provider ollama --swarm --tui
# a real target, cloud model for max quality
export PENTESTSWARM_ORCHESTRATOR_API_KEY=your-key-here
pentestswarm scan <authorized-target> --scope <target> --swarm --follow
# prove findings, don't just report them: confirm each high-stakes finding
# against the live target with a non-LLM oracle (OOB / differential / timing)
pentestswarm scan <authorized-target> --scope <target> --swarm --verify
New here? pentestswarm demo plays the whole campaign offline — that's the GIF above. Running inside GitHub Actions? See deploy/github-action/example-workflow.yml.
Use cases
Each use case is a one-command playbook — a curated swarm configuration for a
specific job. List them with pentestswarm playbook list; run one with
pentestswarm playbook run <name> --target <t> --scope <scope>.
| Use case | What the swarm does | Playbook | Example |
|---|
| Bug bounty — first blood | race a fresh target: recon → classify → exploit → proof, breadth-first | bug-bounty | pentestswarm playbook run bug-bounty --target https://app.example.com --scope example.com |
| External ASM | enumerate a whole external surface and prove the exploitable edges | external-asm | pentestswarm playbook run external-asm --target example.com --scope example.com |
| CI/CD gate | fast, scriptable scan with SARIF output for a pipeline | ci-cd-security | pentestswarm playbook run ci-cd-security --target $STAGING_URL --scope $STAGING_HOST |
| API security (BOLA/IDOR/BFLA) | object-level & function-level authz abuse across an API | api-security | pentestswarm playbook run api-security --target https://api.example.com --scope api.example.com |
| GraphQL audit | introspection, field-level authz, injection, batching abuse | graphql-audit | pentestswarm playbook run graphql-audit --target https://api.example.com/graphql --scope api.example.com |
| OWASP Top 10 | broad web-app coverage mapped to the Top 10 | |
Full walkthroughs with sample output live in the
documentation and the
coverage catalog.
How the swarm works
Key behaviours:
- Agents are independent. Any one of them can be removed, replaced, or added without rewiring the others.
- Pheromones decay per-finding-type. A
PORT_OPEN stays hot for hours; a SESSION for minutes. Config-driven half-lives.
- Scope is enforced at the tool layer and again at the executor. Defence in depth —
--scope is not bypassable.
- Cleanup is always registered before execution. SIGINT, crashes, and budget exhaustion all trigger reverse-order cleanup. See
internal/pipeline/cleanup_memory.go and cleanup.go.
- Prompt caching on Claude cuts cost and latency on repeated system prompts (enabled by default for recon + classifier).
Comparison
How we position vs. the rest of the ecosystem. We'll ship real benchmark numbers in a future release (see the benchmarks roadmap).
| Tool | Open / self-host | Architecture | Executes vs. suggests | Memory | Tools wired | MCP | Swarm? |
|---|
| Pentest Swarm AI | ✅ open, self-hosted | Stigmergic blackboard | Executes | pgvector + pheromones | 8 ProjectDiscovery + nmap; sqlmap / Burp MCP / Metasploit in roadmap | Yes | ✅ real |
| XBOW | ❌ closed SaaS | Autonomous agent (hosted) | Executes | Hosted | Managed | No public API | No |
| PentestGPT | ✅ open | Single-agent ReAct | Suggests | None | None native | No | No |
| HackingBuddyGPT | ✅ open | Single-agent | Executes | Run logs | Shell passthrough | No | No |
| PentAGI | ✅ open | 4 agents + planner | Executes | pgvector | 40+ via MCP/shell | Partial | Pipeline |
| Shannon | ✅ open | White-box + browser | Executes | Session state | Browser DOM | No | Pipeline |
| HexStrike | ✅ open | MCP tool wrapper | Delegates to client LLM | None (stateless) | 150+ via MCP | Yes | No |
| Pentest-R1 | ✅ open (model) | RL-tuned LLM | Executes | Trajectory | CTF-scope | No | No |
If any entry here is wrong or out of date, please open a PR — we want this table to stay honest.
Feature status
Honesty labels: stable means shipped + tested, beta means works but rough edges, alpha means experimental, planned means in the roadmap.
| Feature | Status | Notes |
|---|
| Sequential 5-phase runner | stable | Default mode; battle-tested core |
| Stigmergic swarm scheduler | alpha | --swarm flag; memory-backed blackboard wired |
| ProjectDiscovery toolchain | stable | subfinder, httpx, nuclei, naabu, katana, dnsx, gau |
nmap adapter | stable | XML parsed; scope-validated |
| Cleanup registry | stable | Always runs on SIGINT / exit / budget-cancel |
| Claude prompt caching | stable | Enabled for recon + classifier by default |
--strict LLM mode | stable | Promotes LLM errors to fatal |
| CVSS v3.1 scoring | stable | FIRST spec |
| Jev false-positive filter | beta | --jev; optional second-opinion pass via TypeSafe's Jev — off by default, fails open. Docs |
| Jev adaptive attack-path scoring | beta | --jev-adaptive; the swarm scores its candidate attack strategies with Jev in real time and pursues the best first (graded pheromone). Off by default, fails open. |
| External verification oracles | beta | --verify; confirm a finding against the live target with a non-LLM signal — out-of-band callback, differential response, or timing — and trust that verdict over the model. Proven findings get a VERIFIED badge and a 1.0 attack-graph edge. Scope-guarded; off by default; ignored under --dry-run/--safe-mode. Docs |
| Exploit chains | beta | chain run/list/update; a growing library of named, CVE-tied exploit chains (SSRF→RCE, auth-bypass→RCE…) — fingerprint + safe verify + report. Lean binary + on-demand updates. |
CLI
pentestswarm run # ⭐ Interactive TUI — pick options + target, no flags
pentestswarm scan <target> --scope <scope> --swarm # Scriptable: stigmergic swarm scheduler
pentestswarm scan <target> --scope <scope> --tui # Full-screen live TUI (charts + swarm topology)
pentestswarm scan --lab --lab-target crapi # Attack a bundled vulnerable lab, no setup
pentestswarm playbook run <name> --target <t> # Run a community playbook
pentestswarm doctor # System health check
pentestswarm install-tools # Fetch the recon/exploit toolchain
pentestswarm mcp serve # MCP server for Claude/Cursor
pentestswarm serve # Start API server + dashboard
pentestswarm run is the front door — an interactive launcher (provider + key entry, target or bundled lab, live-view choice, readiness checks) so you never have to remember flags. The scan form stays fully scriptable for CI.
LLM Providers
All agents inherit from a single provider config. Set one key, the entire swarm works.
See the model comparison and recommendations to choose by quality, cost, privacy, refusal behaviour, and setup effort.
Bring your own model — we're the harness, not the model. A new wave of open models is topping the cyber-offense benchmarks — GLM (5.3), Qwen (3.x), DeepSeek, and security-tuned fine-tunes like Pentest-R1. Pentest Swarm turns any of them — or a frontier model, or a fully-local one — into an operating pentester: real tools, swarm coordination, scope enforcement, and evidence-backed reports. The model does the reasoning; the swarm does the work.
Together AI is first-class — pick together (in pentestswarm run or --provider together) and just add your key; the endpoint is handled for you. Any other OpenAI-API-compatible endpoint (OpenAI, DeepSeek, Groq, …) works via openai + a base-URL, plus first-party Gemini. The only hard requirement is native tool/function calling, which GLM, Qwen, and DeepSeek all support.
| Provider | provider: | Setup | Privacy | Best for |
|---|
| Claude (default) | claude | export PENTESTSWARM_ORCHESTRATOR_API_KEY=... | Cloud | Best quality, zero setup, prompt caching |
| Together AI | together | Just set the key — endpoint auto-configured | Cloud | Open cyber-benchmark leaders: GLM zai-org/GLM-5.3, Qwen Qwen/..., DeepSeek, Kimi |
| OpenAI-compatible | openai | Set key + the vendor's /v1 endpoint | Cloud | OpenAI, DeepSeek, Groq, or any Chat-Completions API |
| Gemini | gemini | export PENTESTSWARM_ORCHESTRATOR_API_KEY=AIza... | Cloud | Large context, free tier |
| Ollama | ollama | Install Ollama + pull models | 100% local | Full privacy, air-gapped (GLM / Qwen builds available) |
| LM Studio | lmstudio | Load model, enable server | 100% local | GUI model management |
| OrcaRouter | orcarouter | export PENTESTSWARM_ORCHESTRATOR_API_KEY=sk-orca-... | Cloud | One endpoint for Claude/GPT + other frontier models, gateway-level security |
Together AI example — run the whole swarm on GLM 5.3:
orchestrator:
provider: "together" # first-class — endpoint defaults to Together's API
model: "zai-org/GLM-5.3" # or Qwen/..., deepseek-ai/..., etc. — see together.ai/models
api_key: "" # or export PENTESTSWARM_ORCHESTRATOR_API_KEY
context_window: 128000
Tech Stack
| Component | Technology | Why |
|---|
| Platform | Go 1.24 | Single binary, goroutine concurrency, native security tools |
| CLI | Cobra + bubbletea | Beautiful TUI with multi-panel agent view |
| LLM | Claude / Together AI (GLM · Qwen · DeepSeek) / Gemini / OrcaRouter / Ollama / LM Studio | Best quality cloud + open cyber-bench leaders + full privacy local |
| Security Tools | subfinder · httpx · nuclei · naabu · katana · dnsx · gau · nmap | ProjectDiscovery Go libs + nmap subprocess |
| Blackboard | Postgres 16 + pgvector | Transactional writes, vector similarity, pheromone decay in SQL |
| Cache | Redis 7 | Rate limiting, session state |
| Dashboard | Next.js 15 + shadcn/ui + tremor | Dark-first, chart-heavy |
| MCP | JSON-RPC stdio | Claude Desktop + Cursor integration |
Development
git clone https://github.com/Armur-Ai/Pentest-Swarm-AI.git
cd Pentest-Swarm-AI
./scripts/setup.sh # Install tools, start Postgres/Redis/Ollama
make build # Compile binary
make test # Run tests
make dev # Hot-reload development
Regenerate the demo GIF after any CLI change:
brew install vhs # one-off
vhs docs/demo-flashy.tape
Roadmap
- Wave 1 (in flight): real swarm architecture (done), dashboard wire-up, Burp MCP
- Wave 2: sqlmap / Metasploit / ZAP adapters, bug-bounty + ASM + CI/CD playbook polish, official GitHub Action in Marketplace
- Wave 3: fine-tuned Pentest-Swarm model (Pentest-R1 recipe), Cybench / AutoPenBench / CVE-Bench numbers, agent-memory poisoning hardening (MINJA / MemoryGraft defences)
🧠 The Adaptive Swarm — headline releases
Today the swarm reacts to findings and runs verified attack playbooks (BOLA/IDOR, mass assignment, NoSQL injection, excessive data exposure) end-to-end. Next, we make it think — three major capabilities, each shipping as its own release:
- ① Runtime reaction to discoveries (planned) — stigmergic emergence: the swarm mines every response for object references (ids, UUIDs, emails), writes them to the blackboard, and other agents react by probing those objects across endpoints. Find one leaked id and the swarm turns it into cross-user BOLA/IDOR attacks nobody scripted.
- ② Self-correcting attacks (planned) — closed-loop replanning: a failed step (401/403/415, an odd response body) feeds back into the planner, which adjusts the request and retries. Attacks heal themselves instead of dead-ending.
- ③ On-demand specialist sub-agents (planned) — the swarm spawns purpose-built agents at runtime (an auth agent to hold a session, a fuzzing agent for a discovered parameter, a chain-builder for a specific API) and tears them down when done.
Together these turn a reactive swarm into an adaptive one — attack surface it has never seen, handled without anyone writing a plan. Generalization beyond curated targets and full LLM-driven chaining ride on top of these.
Follow the GitHub Project board for live status.
Why "Swarm"?
Single agents are tools. Pipelines dressed up as agents are slightly fancier tools. A swarm is different: agents share an environment, each agent's writes influence other agents' behaviour, and the useful work is emergent rather than prescribed. That's what lets a swarm handle a 1,000-subdomain target without anyone writing a plan for it.
One agent is a tool. A swarm is a platform.
Our Mission
Offense is going autonomous — the defense of it should be open. XBOW proved AI
can top the bug-bounty leaderboards; it did so as a closed, hosted SaaS. We
believe the capability to find and prove vulnerabilities at machine speed
shouldn't be locked behind one vendor's cloud and price — it should be
open-source, self-hosted, and inspectable, so defenders, researchers, and
independent testers can run it on their own infrastructure and their own models.
So our mission is to build the first real open-source pentest swarm in the
open: a harness that turns any model — frontier or fully-local — into an
operating pentester, with the safety rails (fail-closed scope, egress guard,
an authorization gate, a full audit trail) that keep autonomous offense
accountable. Powerful real-world-harm capabilities stay gated behind verified,
contracted operators — on purpose.
If a real open-source swarm is something you want to exist, help us build
it.
Community
We're building the first real open-source pentest swarm in the open — come build it with us. In Discord you can share findings, request a tool adapter, argue about stigmergy and pheromone decay, get help running your first scan, or grab a good first issue and ship a PR. Researchers, red-teamers, and the AI-security-curious all welcome.
⭐ If a real open-source swarm is something you want to exist, drop a star. Star velocity is the fuel that keeps this shipping — it's the single biggest thing you can do in ten seconds.
Who's using Pentest Swarm?
Running Pentest Swarm — internally, on client engagements, in CI, or embedded in your own workflow? Add your org to ADOPTERS.md with a quick PR — it helps others trust the project and helps us prioritize what to build. Not ready to be listed publicly? A hello in Discord still helps.
Contributors
Every PR, issue, tool adapter, and playbook makes the swarm better. Thank you to
everyone building this in the open. 💛
New contributor? Grab a good first issue and read CONTRIBUTING.md to get started.
Editions — Free · Commercial · Enterprise
Pentest Swarm AI is free and open source (AGPL-3.0), and always will be. The open-source edition is a fully capable, defensive-grade autonomous API/web-app pentester — it exploits and chains attacks (BOLA/IDOR, JWT, SSRF, SQLi, GraphQL, BFLA…), builds and scores attack paths, ships a large library of attack-path + CVE chains, and includes every safety guardrail (fail-closed scope, egress guard, the authorization gate, and a full audit trail).
More aggressive offensive capabilities are deliberately not in the open-source tool. Post-exploitation and host takeover, credential testing at scale, red-team detection-evasion tradecraft, and similar high-impact capabilities are reserved for the Commercial and Enterprise tiers — and are gated behind identity verification and a signed authorized-use agreement. This is intentional, not a paywall trick: capabilities that are trivially abused for real-world harm should not be handed to anonymous users. Access requires completing verification so there is an accountable, contracted operator behind every powerful action.
| Free / Community (OSS) | Commercial (verified) | Enterprise |
|---|
| Autonomous web/API exploitation + chaining | ✅ | ✅ | ✅ |
| All safety guardrails (scope · egress · authorization · audit) | ✅ | ✅ | ✅ |
| Multi-target batch · context ingestion · local memory · coverage catalog | ✅ | ✅ | ✅ |
| Advanced offensive reach — post-exploitation & host takeover, credential testing at scale, red-team evasion tradecraft, and more | — | ✅ (verified + contracted) | ✅ |
| Fleet & platform — parallel fleet dashboard, RBAC/SSO, scheduling, managed/hosted | — | — | ✅ |
| Compliance & assurance — signed/tamper-evident audit, SIEM/SOAR, PCI/SOC2/HIPAA reporting | — | — | ✅ |
| Live chain feed + integrations (Jira/ServiceNow, Burp, CI/CD, cloud, API/webhooks) | large, each release | ✅ live feed | ✅ live feed + full integrations |
| Support | community / Discord | priority | ✅ SLA · onboarding · custom chains · managed service |
Want Commercial or Enterprise access? It requires identity verification and a signed authorized-use agreement — a deliberate safety gate so powerful offensive capability only reaches accountable, authorized operators. Reach out and we'll walk you through verification.
📧 [email protected] — tell us your setup and what you're trying to do, and we'll get you set up.
Legal & Credits
Legal Disclaimer: Pentest Swarm AI is designed exclusively for authorized security testing, bug bounty programs, CTF competitions, and educational research. You must obtain explicit written permission from the target system owner before running any scan. Unauthorized access to computer systems is illegal under the Computer Fraud and Abuse Act (CFAA), the Computer Misuse Act, and equivalent laws worldwide. The authors and contributors of this project accept no liability for misuse, damage, or any illegal activity conducted with this tool. By using this software, you agree that you are solely responsible for ensuring your use complies with all applicable laws and regulations. Do not use this tool against systems you do not own or have explicit authorization to test. The tool enforces scope fail-closed and requires you to affirm authorization before attacking a real target — see the full Responsible Use Policy.
Credits & Inspiration
This project stands on the shoulders of giants. We credit and thank these projects for pioneering AI-powered offensive security:
- PentestGPT — the OG that proved LLMs can pentest
- PentAGI — fully autonomous agent architecture
- Strix — AI hackers that find and fix vulns
- CAI — cybersecurity AI framework, 3600x faster than humans
- HackingBuddyGPT — LLM hacking in 50 lines of code
- Shannon — white-box AI pentester
- BlacksmithAI — multi-agent pentest framework
- PentestAgent — black-box AI security testing
Their open-source contributions made tools like this possible.
Licensing and commercial editions
Pentest Swarm AI follows an open-core model. The Community Edition is licensed
under GNU AGPL-3.0-only; see LICENSE. Commercial use of the Community Edition is
permitted subject to that license. Armur AI also offers commercial, enterprise,
and hosted products, which may include separately licensed components.
For the planned separate-control-plane product, the Community Edition will run
as a separately executed, pinned, unmodified public release under AGPL. This
product architecture does not depend on proprietary relicensing of existing
community contributions.
A signed contributor license agreement may grant Armur AI additional rights,
including rights to distribute covered contributions under proprietary terms,
while contributors retain copyright. Earlier contributions are not covered by
a new CLA merely because the policy is introduced. See CONTRIBUTING.md for the
applicable policy and effective date.
Distribution and modification can create notice and corresponding-source
obligations. AGPL section 13 addresses remote interaction with modified
versions. The license text controls; the price charged for a product or service
does not determine its obligations.
Built by Armur AI.
