Agentic penetration testing

Agentic penetration testing built on audits going back to 2006

Every Security Brigade audit since 2006 went into building it. You authorise the scope. B-52 maps the application, writes the test cases it needs, works through them, and reports only what it actually exploited. Every finding carries the request, the response and the steps to run it again.

Request a Scoping Call Walk through a completed run

One run, end to end

Discovery

What is there, inside the scope you signed off
Planning

A mindmap, and the test cases this application needs
Scanning

Candidates, none of which is a finding yet
Exploitation

Proved, or dropped. Nothing survives on a guess
Reporting

The report, and the findings in your dashboard
QA

A gate that can send the finished report back

QA can send it back to reporting

The sixth phase runs on the finished report. When it sends one back, reporting runs again.

$500

Entry tier, per scan

Coverage classes

1–3 days

Median to report

One scan covers one application or one target. Sign up, pay by card and run it, or bring us an estate and we will scope it with you.

Turnaround is an observed median for the fully autonomous model, from scope sign-off to report delivery. It is not a service level.

Why it can do this

Where the testing knowledge actually came from

Security Brigade has been CERT-In empanelled since 2008, and every engagement the firm has run since it started in 2006 was worked inside Lemon, the firm’s own assessment platform. The test cases, the vulnerabilities and the threat models stayed with the firm instead of leaving with the auditor who wrote them. That record is what our models were trained on.

The harness

It runs the four practices a Security Brigade auditor runs

Mindmap creation: Before anything is tested, the application is drawn: every entry point, every role, and what each role can reach from where it starts.
Test-case generation: The checks come from what this particular application does, so a payments flow and a document portal are never handed the same list.
Comprehensive JavaScript analysis: Client-side code is read for the routes and parameters the interface never links to. An assessment that only follows links stops at the edge of the menu.
Functional flow analysis: Multi-step flows are taken out of order and taken by the wrong role, which is where business-logic defects live and where a scanner has nothing to say.

Autonomy is the easy half of this category to claim, and most of the field already claims it. The half that cannot be bought is human audit practice going back to 2006, written down in a form something else can run.

Autonomy adds two things on top: exhaustiveness and cadence. The generated test set is worked through in full, never sampled down to fit a testing window. The same method reaches every target. And a run happens as often as you release, not as often as a calendar allows.

How a run is put together What an AI pentest actually runs

What that produced

We measured it against our own assessors

B-52 was run in parallel with Security Brigade’s own expert assessment team, on the same targets, and both sets of findings were counted together. B-52 reached 90–95% of that combined set, and it surfaced issues the human team did not.

Comparable coverage, different blind spots. Read the second half of that sentence as the case for the expert-verified model: the platform and the auditor together account for more than either of them does alone, which is exactly why the middle tier exists.

Where the human is

Choose how much human you want on it

All three cover the same eleven classes, so the choice is never about what gets looked at. It is about who checks the work, and about what you have to be able to file afterwards.

The three delivery models and their regulatory consequence
Delivery model	Where the human is	What you can file it as
Fully autonomous	You authorise scope and targets. Nobody acts after that: the platform tests, chains, validates and delivers.	Not signable under Security Brigade’s CERT-In empanelment.
Autonomous, expert verified	A senior auditor checks every finding before it reaches you.	Signable under Security Brigade’s CERT-In empanelment.
Human led	A senior auditor runs the engagement with B-52 underneath.	Signable under Security Brigade’s CERT-In empanelment.

Key

An empanelled auditor is in the engagement, so the output can be filed
No empanelled auditor in the engagement

Three actions stop and wait for you

Autonomy is bounded by written approval, and the boundary is the same in every model:

Destructive or state-changing actions against a production system.
Persistence, implants and lateral movement past the host B-52 first landed on.
Anything that touches live credentials or real customer data.

Everything else runs without asking. Once persistence is approved for a scope, B-52 establishes it. No point in the run hands that part to a person.

Every finding arrives with its exploit

The request, the response, and the steps that reproduce it, in all three delivery models. If your engineers cannot reproduce a finding, they still spend a week chasing it.

In the fully autonomous model, findings additionally pass an independent automated cross-check before they are reported. The exploit is what makes a finding verifiable; the cross-check is a second gate on top of it.

A regulated filing has to be signed by an empanelled auditor, so route it to the expert-verified or human-led model. CERT-In empanelment attaches to Security Brigade, and it is a condition of the testing itself.

One finding, in full

This is what arrives, for every finding

A run against an Indian co-operative bank found a challenge endpoint that answered without a session. Nothing in the response is malformed, so a scanner reports nothing. What follows is the whole of what your engineers receive.

Co-operative bank · India

The CAPTCHA endpoint returned the answer to its own challenge

Severity: High
CVSS: 7.5
CWE: CWE-200
OWASP: A07:2025

Request

GET /rekyc/GetChallengeText HTTP/1.1
Host: digitalservices.northfield-bank.example
Accept: */*

Response

HTTP/1.1 200 OK
Content-Type: text/plain; charset=utf-8
Content-Length: 5

q4Xbe

No session cookie was sent and none was required. The body is the solution to the CAPTCHA the login form is about to present.

Reproduction

Request the endpoint with no cookie, no token and no prior page load.
Read the response body. It is the answer to the next challenge issued.
Submit it with an automated attempt against the re-KYC login, the administrative login, or the customer verification flow.
Repeat. The challenge costs an attacker nothing, so every flow it guards is unrated.

At the QA gate

Raised from medium to high. The finding was first rated on disclosure alone; the QA gate rated it on consequence, because the endpoint needs no session and the remaining work to exploit it is trivial.

Where you are coming from

Most buyers arrive for one of three reasons

The platform is one platform. What changes between these three is the cadence, the delivery model and what the output has to survive afterwards.

The three triggers this platform is bought on, and where each one goes next
What started it	What that actually needs	Where to start
You ship faster than the testing calendar	A run inside the pipeline that fails a build on a severity threshold you set. Not a report six weeks after the release it was about.	See how a standing cadence runs
The estate is larger than the engagement	Coverage of the applications that never make the annual scope. Not the unimportant ones. The ones there were never enough auditor hours to reach.	See how a run works
A regulator is asking	Output an empanelled auditor can sign, with the findings mapped to the controls you file against. That is the expert-verified or human-led model, and not the fully autonomous one.	Request a scoping call

The three triggers this platform is bought on, and where each one goes next

You ship faster than the testing calendar

What that actually needs: A run inside the pipeline that fails a build on a severity threshold you set. Not a report six weeks after the release it was about.
Where to start: See how a standing cadence runs

The estate is larger than the engagement

What that actually needs: Coverage of the applications that never make the annual scope. Not the unimportant ones. The ones there were never enough auditor hours to reach.
Where to start: See how a run works

A regulator is asking

What that actually needs: Output an empanelled auditor can sign, with the findings mapped to the controls you file against. That is the expert-verified or human-led model, and not the fully autonomous one.
Where to start: Request a scoping call

What it can be pointed at

Eleven coverage classes, one platform

Every class below is performed by the platform, in every delivery model. Mobile is the one to read first if you have been told autonomous testing stops at the browser: B-52 decompiles and analyses APK and IPA binaries itself.

Web applications

Application
Mobile applications

Application
APIs

Application
Thick client and desktop

Application
External network and perimeter

Infrastructure
Internal network

Infrastructure
Cloud configuration

Infrastructure
Active Directory and identity

Infrastructure
Social engineering

Beyond the estate
Source-code review

Beyond the estate
AI and LLM applications

Beyond the estate

Physical security, hardware and wireless testing are out of scope, and that is the only exclusion.

The entry tier is $500

That is one scan, and one scan is one application or one target. Card payment works without a sales conversation. If your estate is bigger than one target, or the output has to be signed, book a scoping call instead.

Request a Scoping Call Who builds it