Most security teams have lived with the same report. Two hundred findings, sorted by a severity nobody agrees with, half of them already known, a third of them not real. Someone spends a week working out which is which, and by then the release has shipped anyway.
False positives are not a cosmetic problem. Each one teaches the reader to trust the next finding a little less. Eventually the real one gets the same shrug as the rest.
The rule
Inside AIpento, a finding from a live test has to reproduce before you see it. An AI model never sends a request to your app: it proposes which test to run, and a separate component written in strict code - the Prover - is the only thing that runs it. The Prover records the exact request it sent and the response that came back. If that exchange shows the weakness, the finding is kept with its request and response attached. If the test returns nothing, the finding is withdrawn.
Withdrawn findings do not appear in your report with a lower score. They do not appear at all.
Why the judge is not an AI
AIpento runs on Quorum, a framework of seven agents with separate jobs. Three of them are AI models: they map your app, plan the tests and look for ways to chain small weaknesses into a real attack. Those are jobs where judgement and imagination help.
The other four are strict code. One runs the order of work and checks each agent's output before it moves on. One, the Prover, is the only part of AIpento that ever sends a request to your app, and it only sends requests inside the scope you authorised. One rules on the evidence. One keeps the record.
We built it this way on purpose. Agents propose; strict code decides. A model is good at suggesting where a weakness might be. It should not be the thing that decides whether a weakness is real, because then a confident mistake becomes a finding.
What you get instead of a pile
- Every finding carries its proof: the request, the response, and what the response shows.
- Findings that only make sense together are linked into one path, the way an attacker would use them.
- Some questions need a person: whether an endpoint is meant to be public, for example. Those are marked "needs your confirmation" and never decide the verdict on their own.
- The release gets one answer: Ship, or Fix first, backed by an Evidence Pack you can hand to an auditor.
In the product it is not dramatic: a row in an evidence log that says what was sent, what came back, and which of the two it was - kept, or withdrawn.