Skip to content
Offensive security platform

Autonomous Red Teaming,
Enterprise Scale.

An orchestrator deploys its specialists against your target, chains what they find, and hands every candidate to a validator with no prior context. What survives reaches you as a submission-ready report.

Request a briefingSee how it hunts
Diffixdiffixlive

Confirmed findings

0

186 candidates · 172 rejected cold

+4 today

By severity

  • Critical3
  • High6
  • Medium4
  • Low1

Mean cold validation

4m 12s

92.5% of candidates rejected

Squad activity

  • Recon1,284 runs0 hits
  • Exploit412 runs9 hits
  • Coverage268 runs3 hits
  • Chain47 runs2 hits
Walkthrough
Overview

01 — The problem

Research is bounded by researcher-hours.

A scanner finds what it has a signature for. A researcher finds what they have time for. Neither scales to the surface of a modern application, and neither chains two medium findings into the critical that was always there.

An engagement should be limited by the target’s attack surface, not by how many hours one person can give it.

02 — The roster

One orchestrator, seven squads.

Each agent is narrow on purpose: its own prompt, its own toolset, its own golden image. Narrow agents keep context clean and stay auditable alone.

O

Orchestrator

Holds engagement state, delegates, and re-plans as findings land. Never runs a fixed script.

R

Recon

Surface

  • Mapper

    Routes, parameters, auth boundaries

  • Operator

    Browser-native, CDP pool, live session

A

Analysis

Intelligence

  • Dependency

    Manifests, lockfiles, transitive trees

  • SCA / CVE

    Components against known advisories

  • Minified JS

    Bundle unpacking, source maps, dead routes

  • Secrets

    Keys, tokens and endpoints left in client artefacts

  • Research

    Exploit-DB, OSINT, disclosed bounty reports

E

Exploit

Pressure

  • Exploit

    Primitive construction

  • XSS

    Injection contexts, escape chains

  • RCE

    Command and code execution

  • SSRF

    Internal reachability

  • API security

    Schema and object-level authz

  • Business logic

    Workflow and state abuse

C

Coverage

Completeness

  • OWASP

    Top-10 conformance sweep

  • WSTG

    Testing-guide checklist

  • Mobile

    iOS and Android targets

X

Chain

Composition

  • ATT&CK

    Technique mapping across confirmed findings

  • Kill chain

    Recon to foothold to privilege escalation

  • Pivot graph

    Reachability between findings and internal hosts

  • Escalation

    Promotes chained mediums to a single critical

V

Verify

Proof

  • Validator

    Zero prior context. Receives the claim and the target, nothing else, and reproduces the bug cold or rejects it

P

Report

Delivery

  • Reporter

    Writes the submission draft in the programme’s own field spine, grounded in the evidence store

03 — Architecture

Hunt context never crosses the validator boundary.

A system that validates its own findings will confirm its own hallucinations. The reasoning that produced a candidate is exactly the reasoning that will rationalise it, so Diffix severs the connection.

Inside the hunt

ReconAnalysisExploitCoverageChain

Shared transcript, shared reasoning, shared state. Everything these agents learn stays on this side.

Context stops here

Across the boundary

Validator

Receives the claim and the target. No transcript, no reasoning, no path.

Reporter

Writes only what the evidence store holds, each claim with its artefact.

It reproduces the bug cold, or the finding never reaches you. False positives are the tax every scanner charges. This is how we refuse it.

04 — The report

Report

Written for the person who has to reproduce it.

Diffix does not hand you a vibe. It hands you a file: severity, CWE, affected object, session, requests, and a replay the validator already ran cold. You verify. You submit. The world gets one bug safer.

  • Validator signature on every confirmed finding
  • Human gate before any disclosure or ticket

DIFFIX / FINDING / 0x4F2A

CRITICAL · CONFIRMED

Broken object-level authorization

CWE-639
GET /v2/invoices/{id}
tenant A token → tenant B invoice

reproduction

1. authenticate as org-admin of tenant A
2. replace path id with tenant B invoice
3. 200 + full PII in body

validator: reproduced from empty context

human: pending

05 — The controls

Autonomy with a leash held in code.

An agent that drives full exploitation has to be bounded by something stronger than instructions. These four are foundational, not features.

  1. 01

    Authorization before autonomy

    No agent acts outside a signed engagement. Scope lives in the control plane and is checked at the tool boundary.

  2. 02

    Tool output is data, never instruction

    Agents read attacker-controlled content all day. Nothing reaches a reasoning prompt without crossing a trust boundary.

  3. 03

    Nothing asserted without evidence

    The reporter may only state what the evidence store holds, with a citation. No artefact, no claim.

  4. 04

    The last call is human

    You comment, steer and clarify while the hunt runs. Nothing is submitted to a programme without you reading it.

Early access

Diffix is an early research system.

We are working with a small number of researchers and security teams before general access. If you run a programme, or hunt them on hardened targets, we would like to talk to you.

Request a briefingfounders@diffix.ai
  • No claim we cannot demonstrate.
  • No finding we have not reproduced.
  • No engagement without signed scope.