AI-Driven Pentest Platform
Autonomous, Logic-Aware Penetration Testing
A design for AI agents that test an application knowing its roles and workflows, prove a finding by demonstrating it safely, and stop for human approval before anything changes state. The judgement stays with the expert.
This page describes how the platform is meant to work. It is not built yet, so there is nothing to buy, install or demo today. Penetration testing itself is delivered now, by our experts, as a service.
An annual penetration test covers a few weeks of the year, and most of its hours go on mechanical cases: the same checks for every role and every workflow. The expensive findings sit in business logic and need an expert’s judgement, and that judgement is what runs out first.
The AI-Driven Pentest Platform is our design for handing the mechanical part of that work to AI agents, inside strict limits. The agents would work from the application profile kept on our platform, run only the cases an expert has marked as eligible, prove each finding by demonstrating it safely, and stop for human approval before any step that changes state. What they find would join the same finding lifecycle as the experts’ work.
It is the last capability on our platform roadmap and it is not built yet. We publish the design so you can judge the approach and tell us whether it fits a problem you have. Until then, penetration testing is delivered by our experts, as a service.
How the Work Is Divided
Machines run what machines do well, and the expert judges the rest
A Look at a Run
The agent at work, the steps held for approval, and what was actually proved at the end
An illustration of the interface with sample data. Switch screens with the tabs above.
What It Is Designed to Do
What would make autonomous testing trustworthy, all of it at the design stage
What It Would Test
Anywhere there are logins, roles and workflows
Use Cases
Where expert hours are scarce and applications are many
How It Would Work
From authorization to a reviewed report
How It Would Be Offered
Our plan once it is built: with an engagement, installed with you, or on a schedule
Why This Platform
How it differs from an automated scanner
Frequently Asked Questions
What teams ask before letting an agent test anything
It is at the design stage. This page describes how the platform is intended to work, not software you can install and run today — it is the last capability on our platform roadmap and is not yet built. We are sharing the design openly; if it fits a problem you have, talk to us about an early engagement rather than a download.
No, and a vendor who says otherwise is selling you something. It takes the mechanical subset of a methodology — the cases a person marked agent-eligible — so expert hours go to the logic and the judgement no machine handles. The report says which parts the machine covered.
It can, with limits you set, but a staging environment that mirrors production is the usual choice for anything beyond read-only cases. Where production is the only option, state-changing steps wait for approval and the request rate stays inside what you agreed.
Four things: sandboxed execution, an outbound allow-list so it cannot wander out of scope, request-rate and concurrency limits you set, and human approval before any step that writes, deletes, pays or sends a message.
A candidate becomes a finding only when it has been demonstrated safely and an expert has agreed. Things that cannot be demonstrated are listed as uncertain, separately, so nobody confuses a hunch with a result.
As designed, nothing that identifies you: requests would pass through Hafez Fabric, our governed AI gateway, which can mask sensitive values before anything reaches an external model. Installed inside your network and pointed at local models, no request would leave at all.
A scanner matches patterns against endpoints. This is given the application profile — roles, workflows, rules — and works through a methodology built for that application, including multi-step sequences across roles that a pattern matcher cannot express.
Authorization and scope, a test environment, accounts for each role, and whatever you know about the application — its workflows and the rules that must hold. If it has been assessed with us before, the profile is already there.
A person you nominate, and not the person who started the run. Approvals are recorded with who granted them and when, which matters when an auditor asks how an automated test was controlled.
Request and token limits per engagement, enforced by Hafez Fabric, the gateway the agents would run through, and visible while a run is going rather than at the end of the month. A run that hits its limit would stop and tell you.
Confirmed findings with evidence and reproduction steps, the uncertain candidates listed separately, what the run covered and what it did not, and a replayable record of every request and decision the agent made.
That is the case it is designed for. Given accounts for each role, it would work across them: what one role can do to another’s data, which step can be skipped, what happens when a sequence is repeated or interleaved.
It would share VAaaS’s methodology, application profile and finding lifecycle. It is the part of the methodology a machine could carry, so that an assessment could cover more ground in the same hours rather than simply cost less.
Related Services
Complementary services that might be useful for you
Related insights
What we have written about this service
Other Platforms
The rest of the HafezSecure platform range
Vulnerability Assessment as a Service
Attack Surface Management
Cyber Asset Inventory