Skip to content

Seven agents · one shared brain

FIND THE FLAW. PROVE THE FIX.

A repair counts only when independent, executed tests agree.

Go to projectsHow It Works
Illustration · not a recorded runVerifyreplay · 403 Forbidden · exploit blocked 2/2 · fix verified
  1. 01Understand
  2. 02Review
  3. 03Attack Safely
  4. 04Diagnose
  5. 05Repair
  6. 06Verify
  7. 07Monitor

The problem

Why every change needs checking

Code changes every day.

Developers and AI agents ship new behaviour with every merge, and each one can change what the software does.

Reviews see a diff. Scans see a snapshot.

A reviewer reads the lines that changed. A scanner checks a moment in time. Intent and real exploitability go unchecked.

EVERY CHANGE. CHECKED.

Every change is tested against what the software is supposed to do and probed like an attacker would, then re-checked as code moves.

Illustration · not measured

How it works

Assurance that runs
on every change.

ARI learns how the software is meant to behave, tests that behaviour safely, repairs only with approval, and keeps checking as the code moves.

  1. Illustration · not a recorded run

    Learn what it should do

    Code intelligence indexes the repository and extracts expectations from specs, docs, tests and history.

    Understand intent

  2. Illustration · not a recorded run

    Test it like an attacker

    The Reviewer writes tests that reproduce defects, and seven security agents probe inside an approved sandbox scope. Every authorization expires: setup requests 30 days, and none can exceed 90 days.

    Review and attack

  3. Illustration · not a recorded run

    Fix it, then prove it

    After a human approves, Patch proposes the smallest fix. Verify replays independent tests on a fresh checkout.

    Repair and verify

  4. Illustration · not a recorded run

    Keep it secure

    Signed GitHub events re-run the assessment as code changes, and every finding feeds the next test.

    Monitor and retest

One finding, start to proof

Four harmless steps.
One verified fix.

  1. Map the surfaceScout

    Scout maps endpoints, data flows and trust boundaries inside the approved scope.

  2. Chain the evidenceProbe · Trace

    Probe exercises each harmless-looking step; Trace links them into one exploitable path.

  3. Find the root causeDiagnose

    Diagnose pins the defect to source spans and explains why it happens.

  4. Fix and prove itPatch · Verify

    Patch proposes a fix after approval; Verify replays the exploit and the suite on a fresh checkout.

Illustration · not a recorded run

root cause · create_refundverified · replay 2 / 2

Proof, not promises

The record is checkable.

No model output counts as proof, including ours. In the console, every status comes from saved records.

01Design

7

Separate security agents, each its own tool loop with enforced permissions.

02Rule

2×

A failing test only counts if it fails the same way twice.

03Rule

Never

Remediation approval stays with a different human.