SwarmPT — Human-Led AI Offensive Security

AI Penetration Testing
Swarm scale. Human judgement.

A real attack on your environment, not a scan of it. A swarm of AI agents finds the genuine entry points and works them in parallel — exploitation, business logic abuse, original zero-day research — with an experienced penetration tester driving the whole engagement and pushing the swarm well past where it would stop on its own.

What is AI penetration testing?

AI penetration testing uses autonomous AI agents to carry out the work of a penetration test — reconnaissance, vulnerability discovery, exploitation and reporting — at a scale and speed no human team can match by hand. Instead of one or two testers working sequentially through a time budget, dozens of specialised agents work the target in parallel.

The distinction that actually matters is who is operating the swarm. Agents left to run a playbook on their own will find what a playbook finds. Every Security Arsenal engagement is SwarmPT: an experienced penetration tester at the controls for the duration, reading what comes back, redirecting agents onto whatever looks promising, and pushing the swarm well past where it would stop by itself.

That direction is not a quality-control step added at the end — it is a capability upgrade. A tester who knows what to look for will get the same swarm to attempt things it would never reach alone, and that is consistently where the findings that matter come from. It is also why we sell one engagement at one price rather than a cheaper tier with nobody at the controls.

Measured, not claimed

Twelve live websites. Every one behind a firewall. We got into all twelve.

We ran an ordinary vulnerability scan first, as a control. The firewall detected it and blocked us — which is how we know the protection was live and enforcing. It never saw the engagement that followed, because nothing has a signature for an exploit written while the test is running.

12
Live sites tested
8
Critical findings
39 min
Median per target
3 min
To a written report
0
Times the WAF stopped us
See the full results, including what we did not find
SwarmPT

See the attack run

This is a replay of a real engagement. Watch the violet lines— that is the red team lead typing into the swarm while it runs: throttling it when it trips rate limiting, routing it through the proxy pool when a WAF starts fingerprinting, killing a false positive, and catching the flow the agents walked past. That channel is what "human-led" means.

swarmpt — penetration test — coolittackle.com

Agents

swarmpt
Odoo Database Manager Validator
Authentication & Session Tester
Odoo Injection & XSS Tester
│ ├Odoo IDOR & Info Disclosure Validator
│ ├Odoo Cart Business Logic Validator
Information Disclosure Tester
│ ├Info Disclosure Reporting Agent
Odoo Exploit & CVE Hunter
│ ├Odoo Database Manager Reporter
│ ├Host Header Injection Reporter
Odoo 17.0 RCE & CVE Validator
XSS in Shop Search Validator
Contact Form Misconfig Validator
│ ├Signature Bypass Reporter
Subdomain & Infra Tester

Confirmed findings

swarm working…

SwarmPT0.0M tokensrunning…

Inside the swarm

Different attacks need different specialists. Rather than one generalist model doing everything badly, the engagement runs specialised agents in parallel — each with its own role, tooling and success criteria.

Recon Agents

Map the full attack surface in parallel — subdomains, exposed services, APIs (REST, GraphQL, SOAP), JavaScript bundles, secrets in client code, forgotten staging hosts, and cloud assets nobody remembers provisioning.

Exploitation Agents

Actively chase every candidate finding to proof: SQL injection through WAF bypass, XSS in all contexts, SSRF, IDOR/BOLA, auth and session flaws, deserialization, file upload bypass, and privilege escalation chains.

Business Logic Auditors

The class of bug scanners never find. Agents reason about what your application is *for*, then abuse the workflow — price manipulation, race conditions, entitlement bypass, multi-step approval circumvention.

Zero-Day Researchers

On in-scope engagements, dedicated agents perform original vulnerability research against your custom code and the software you depend on — hunting for unknown flaws rather than replaying published CVEs.

Internal & Lateral Movement Agents

Jump host agents deploy to any machine with internal network access — no VPN, no firewall exceptions. From there: credential attacks, Active Directory escalation paths, and realistic lateral movement.

Human Test Lead

Not a reviewer at the end — the operator throughout, talking to the swarm on a live channel. "You are too loud, throttle and rotate the proxy pool." "WAF is fingerprinting you, drop the tooling UA." "That XSS is a false positive, drop it." "You missed the reset flow." Corrections like these are what separate a swarm being run from a swarm being operated.

AI penetration testing vs. scanners vs. traditional pentests

Automated scanners are cheap and shallow. Traditional consultancy engagements are deep but rationed by hours. A human-led AI swarm is the first option that is both.

 Automated scannerTraditional pentestSecurity Arsenal — human-led AI
CoverageKnown signatures onlyWhat one or two testers reach in the time budgetMany attack paths worked at once, not one after another
Business logic flawsNever foundFound if the tester has timeDedicated agents reasoning about intended behaviour
Zero-day / novel bugsNoneRare — depends on the individual researcherDedicated research agents, in scope on request
TurnaroundMinutes, low value2–6 weeks including schedulingUnder an hour per target on our last run — report included
False positivesHigh — you triage themLowLow — nothing is reported that was not exploited, with the code to prove it
Retest after fixesRe-run, re-triageOften billed separatelyIncluded — plus 90 days of blocking while you fix
Repeatable / continuousYes, but shallowAnnual, by budget realityPer release or per quarter — re-running a known scope is cheap for us

How an engagement runs

Same rigour as a traditional engagement — the difference is what happens in step three.

01

Scope & Rules of Engagement

Targets, test windows, excluded systems and escalation contacts are agreed and signed. The signed SOW is the testing authorisation.

02

Environment replication

An isolated sandbox network is built for your engagement alone and your environment replicated inside it, so the swarm can rehearse before anything active touches production.

03

Swarm execution

Specialised agents run recon, exploitation, logic abuse and — where scoped — original zero-day research, all in parallel.

04

Expert direction

The tester steers throughout — redirecting agents, deepening promising leads, correcting whatever went wrong, and pushing the swarm past where a playbook stops. This step is where the findings that matter come from.

05

Report & briefing

Executive summary, full technical report with CVSS scoring and proof-of-concept evidence, compliance mapping, and a live walkthrough with your team.

06

Remediation retest

After you fix, we retest critical and high findings and reissue the report with confirmed closure status. Included, not billed separately.

What the swarm tests

Scoped to what you actually run. Nothing tested that you have not authorised.

Web applications

  • Full OWASP Top 10
  • Auth, session, JWT & OAuth
  • Business logic abuse
  • Client-side secret exposure

APIs

  • REST, GraphQL, SOAP
  • IDOR / BOLA
  • Mass assignment
  • Rate limit & abuse paths

Networks

  • External attack surface
  • Internal lateral movement
  • Active Directory escalation
  • Credential attacks

Cloud

  • IAM misconfiguration
  • Storage bucket exposure
  • Serverless attack surface
  • Container escape paths

Is it safe to point AI at production?

It is the first question every serious buyer asks, and it should be. Autonomy without limits is how testing turns into an incident. Every engagement is bounded before a single packet moves:

Signed Rules of Engagement

Targets, windows, exclusions and destructive-action limits are documented and signed. The SOW is the authorisation — not a checkbox.

Isolated sandbox first

Your environment is replicated in a network built for your engagement alone. No infrastructure is shared between clients.

Encrypted credential handling

Any credentials you provide live in an encrypted vault, are scoped to the engagement, and are destroyed at close-out.

Human kill switch

A named emergency contact and hotline can pause or halt all testing immediately, at any hour.

AI Penetration Testing — Common Questions

Find what a scanner never will

Tell us what you run and what you are worried about. We scope it, price it, and come back within one business day.