Muhkam Shield

We build with AI. And we hand over nothing we have not tried to break ourselves.

For us, AI is a tool in the hands of a team that works like engineers: written specs before code, a test for every change, and break-in attempts on a copy of the system before it reaches you. We call this Muhkam Shield, and every system we deliver goes through it.

Not AI writing code and hoping

We build the way an engineer builds.

Specs before code

Before a single line is written, the system is written down: what it does, who sees what, and where data is kept. You approve it first.

A plan for every change

Every change has a written plan before it starts: what will change, how we check it worked, and how we roll back if something goes wrong.

A test for every part

Every part has tests that run again with every change, so what was fixed before does not break without anyone noticing.

Independent review

Code is reviewed by someone who did not build it, looking for what is wrong before what works. Nothing passes until every test passes.

A backup before any change

Any change to real data is preceded by a backup, and every change has a written way back. No step is irreversible.

Staged launch

The system goes live in stages: a trial copy, then part of the work, then all of it. Nobody gets surprised halfway.

Before we hand over

We try to break it, so nobody else can.

These are all the tests a system goes through before handover. Where a test has a recognised international name, it is written next to it.

Penetration testing 9 tests

Login and sessions

Password guessing, sessions left open, and password recovery.

OWASP A07 · Authentication Failures
Permissions

A user trying to do something outside their role, or open another customer's or branch's data by changing a number in the link.

OWASP A01 · Broken Access Control · IDOR
Injection

Every field a customer can type into is tried with malicious code, in the database and on screen.

OWASP A03 · Injection · SQLi · XSS
Files

Private documents cannot be opened with a guessable link, and every upload is checked for type and size.

OWASP · Unrestricted File Upload
Secrets

No key or password in the code, in the pages or in the AI's instructions.

CWE-798 · Hard-coded Credentials
Server settings

HTTPS, security headers, and ports closed from outside.

OWASP A05 · Security Misconfiguration
Libraries

Known vulnerabilities checked in every library the system uses.

OWASP A06 · Vulnerable Components
Connections to other systems

Messages coming from WhatsApp or payments cannot be forged.

OWASP API Security Top 10
Logs

Every login and every important action is recorded, so any attempt is seen.

OWASP A09 · Logging & Monitoring

AI testing 9 tests

Prompt injection

We try to make it ignore its instructions or follow orders hidden in a customer's message.

OWASP LLM01 · Prompt Injection
Instruction leaks

We ask it to reveal its internal instructions, and check they hold no secrets.

OWASP LLM07 · System Prompt Leakage
Another customer's data

We ask it about another customer by name or number, and it must reveal nothing.

OWASP LLM02 · Sensitive Information Disclosure
Tool permissions

Every tool it has is locked to the conversation in front of it, whatever it is told.

OWASP LLM06 · Excessive Agency
Usage limits

A limit per conversation, so nobody can drain the system with a flood of messages.

OWASP LLM10 · Unbounded Consumption
Loop protection

If another bot starts replying to it, the system detects the loop and stops it.

A pack of real conversations

The insistent customer, the angry one, the voice note, 3 am, two requests in one message. On every change.

3 runs for every sensitive case

On the real model, because an answer can differ from one run to the next.

Silence is not allowed

If a customer waits more than 2 minutes without a reply, the person in charge knows at once, with the reason.

Trial runs and stress 6 tests

Regression tests

Every old test runs again with every change, so what was fixed stays fixed.

Regression Testing
Dry run

The system runs on your real data without sending a single message, and we read everything it would have done.

Dry Run · Shadow Mode
Stress test

A button pressed twice, the connection dropping, WhatsApp stalling, and many users at once.

Stress Testing
Random cases

Money calculations are tried on thousands of random cases, and every screen must show the same number.

Property-based Testing
On a copy, not your data

Every attack or stress test runs on a separate copy. Your real data is never touched.

Staging Environment
A trial week

After launch, a daily review of conversations and numbers with you, with changes once a day.

Pilot · UAT

This really happened

Not theory. These things happened.

3 attacks in 2 minutes

On 23 October 2026 a stranger tried 3 well-known attacks on our AI assistant: SQL injection, a request for its internal instructions, and a request for another customer's data by number. They got nothing, and the conversation was closed.

A hidden permission

The stress test on a café system found that staff could delete tables behind the screen, even though the button was hidden from them. It was closed before anyone used it.

One press too many

The same test found that pressing "close shift" 3 times counted the money 3 times. It was fixed before the first real shift.

200,000 cases

The money calculations in the same system were tried on 200,000 random cases, and the result matched on every screen.

What you receive

The Shield report, in writing.

What was tested

Every test run on your system, by name: login, permissions, injection, files, AI and stress.

What we found

Every issue found, ranked by severity: critical, high, medium, low. With the steps to reproduce it.

What was fixed

Every fix, with a test that failed before the fix and passes after it. Not words, a test that runs.

After handover

Any mistake that appears after handover is logged and gets a new test, so the same kind never repeats in any system.

Questions

Questions about security

Does the AI see my customers' data?

The AI employee sees only the conversation it is answering. Every tool it has is locked to that conversation's data, and its instructions hold no secret or password.

Where does our data go?

Your data is stored on our servers and never shared with anyone. The AI receives only what it needs to reply, over an encrypted connection, and every system is locked to your company's data alone.

Who can access the system?

Only the people you choose. Every user has permissions that match their job, and every action is recorded under the name of who did it.

What if you find a problem after handover?

It is fixed at once, and it gets a test that runs with every change from then on.

Does Shield cost extra?

No. Every system we deliver goes through Shield, and comes with the report.

Have a question about the security of your data? Ask us before anything else.

We answer in detail, and show you the Shield report on your own demo.

Ask about security
Ask about security