▸ pentest-swarm-ai · open_source

Machine-Speed Attacks.Human-Level Reasoning.

Built for your cloud infra. A swarm of autonomous AI agents that chains vulnerabilities — across APIs, web apps, and infrastructure — the way a real attacker would. Open-source so you can audit the reasoning. Deployed against your stack by the team that built it.

A swarm of specialists. One coordinated attack.

Eight specialist agents share a single blackboard — each one recon, auth, web, cloud, IaC, supply chain, network, exploitation. They chain findings across surfaces the way a real adversary would, and the reasoning is reproducible and auditable. No black-box scanners.

pentest-swarm · sample.loglive
$ pentest-swarm spawn --target your.stack
▸ initializing 8 specialist agents
▸ recon 247 endpoints discovered
▸ auth JWT alg=none → bypass confirmed
▸ web IDOR on /api/users · 14,231 records
▸ cloud s3://acme-backups · PUBLIC READ
▸ chain auth → idor → exfil · 3 hops
✓ engagement complete · 47 findings · audit log written
armur-ai · pentest-swarm-ai
0
stars · agpl-3.0 · built in the public
audit the reasoning · every commitgit clone →

▸ engagements

Want it run against your stack?

The team that built the swarm will deploy it against your production-grade environment, validate findings by hand, and hand back a report your engineers will actually use.

Talk to the team

Eight specialists. One shared brain.

Each agent has a single specialty. They share state through a stigmergic blackboard — a finding from one agent immediately becomes input for the next. That's how the swarm reasons across surfaces a single tool would never chain.

recon/01
enumerate attack surface
247 endpoints discovered
auth/02
session + token attacks
JWT alg=none bypass
web/03
OWASP + business logic
IDOR · 14k records
cloud/04
IAM, storage, compute
s3://*-backups public
iac/05
terraform + kubernetes
admin policy attached
supply/06
deps, typosquats, SBOMs
maintainer takeover
network/07
services + segmentation
smb anon · 47 GB
exploit/08
chain + prove impact
auth → idor → exfil
pentest-swarm · engagement.loglive

Hire the team. Use the full arsenal.

Don't run the swarm alone. The people who wrote it will deploy it against your environment, validate every chain by hand, and walk your engineers through the report. Plus the full open-source arsenal we maintain alongside it.

▸ direct engagementaccepting briefs

The team that built the swarm, deployed against your stack.

▸ engagement lifecycle
[01]
Scope
2-3 days
[02]
Deploy
1 day
[03]
Execute
2-4 wks
[04]
Validate
3-5 days
[05]
Report
48h SLA
[06]
Retest
included
▸ what you get
  • Senior operators run the swarm against your production stack
  • Every chained finding hand-validated — zero noise in the report
  • Audit-ready report + reproducible run logs your engineers can re-run
  • One-off engagement or continuous cadence
Brief the team· 10–15 engagements / year capacity

The Pentest Dojo. Train for the post-AI war room.

A free open-source curriculum for AI-assisted offensive security — the first generation of labs that take post-AI threat models seriously. Agentic recon, model-aware testing, prompt-injection chains, swarm orchestration. Vendor-agnostic. Runs on your laptop.

▸ curriculum path
[01]
Fundamentals
1 module
[02]
Core
2 modules
[03]
AI-Specific
2 modules
[04]
Advanced
2 modules
[05]
Capstone
1 module
// WHAT_YOU_LEARNdifficulty ●○○○ → ●●●●
── Fundamentals1 lab
  • LAB_01agentic_recon●○○○
── Core2 labs
  • LAB_02autonomous_exploit●●○○
  • LAB_05agent_trust_boundaries●●○○
── AI-Specific2 labs
  • LAB_03model_aware_testing●●○○
  • LAB_04prompt_injection_chains●●●○
── Advanced2 labs
  • LAB_06swarm_orchestration●●●○
  • LAB_07supply_chain_for_ai●●●○
── Capstone1 lab
  • LAB_08live_red_team_dojo●●●●
dojo.deploy
Deploy the dojo locally· Apache 2.0 · runs on docker compose