01 · Self-Improving · Autonomous · Sovereign

A self-improving offensive security team.

A vertically-integrated fleet of agents that pentests your apps, mines every finding into training data, and finetunes its own next-generation model — continuously.

02 · The Platform

Five products. One autonomous loop.

01
Djini AI
Mobile & API Security
End-to-end discovery and exploitation of advanced bugs across iOS, Android, and their API back-ends — driven by real user interaction, business-logic reasoning, and native-layer analysis. Auto-login, screen detection & reading, chained exploit construction.
02
Research Lab
Browser Exploitation · WebKit · Chrome
Discovers and combines real exploit chains for WebKit and Chrome — building, testing, and validating them autonomously on the device-lab. From primitives to working chain, no analyst in the loop.
03
Finetune Studio
Self-Improving Models · The Brain
Auto-finetunes Djini's models by detecting skill gaps, generating synthetic data from real Research-Lab and device-test outputs, and re-evaluating until the gap closes. The model improves itself.
04
Device Lab
Hardware Foundation · iOS 26 & Android
Our own emulation fleet of jailbroken iPhones (iOS 26.*) and rooted Androids — plus real, selectable devices for every test. The physical substrate that turns AI hypotheses into working exploits.
05
AppSec Portal
Your Security Team, As Agents
A full agent-based security organization — PO, Security VP, mobile engineers, compliance & risk — embedded into your stack with real identities. Agents pick up tickets, message engineers on Teams, and update compliance boards autonomously.
03 · How It Works

A platform that writes its own next test.

Six autonomous stages. Every product is a tool and a data source. Frontier models fill skill gaps. Multiple judges validate every release. Zero humans in the loop.

01 · TEST Tests apps & APIs mobile · web · native targets Djini AI · Research Lab 02 · CAPTURE Generates datasets findings · traces · chains every product, every run 03 · SYNTHESISE Fills skill gaps frontier models · synthetic data targeted · adaptive 04 · TRAIN Releases its own models finetune · version · ship Finetune Studio 05 · VALIDATE Multi-judge eval independent · adversarial 5 judges · zero regression 06 · IMPROVE Suggests code fixes to itself · to your apps PRs · scanner upgrades
04 · The Compounding Loop

Every test makes the next test smarter.

Every product is both a tool and a data source. Research findings become training data. Training data becomes better agents. Better agents find harder bugs. The system never stops — and the moat compounds with every device-hour.

DJINI flywheel ∞ AUTONOMOUS DJINI AI Mobile · API · Native tests & exploits apps RESEARCH LAB WebKit · Chrome builds exploit chains DEVICE LAB iOS · Android · Real validates & runs PoCs FINETUNE STUDIO Train · Eval · Ship closes skill gaps APPSEC PORTAL Orchestrate · Agents your team, as agents
05 · Sovereign by Design

Fully local. Fully yours.

01

On-premise

Every component — orchestrator, agents, training stack, eval harness — deployed entirely inside your environment. Air-gap supported. Zero outbound traffic.

02

Your devices

Jailbroken iPhones, rooted Androids, real selectable handsets — racked in your DC. Unlimited autonomous test-hours, never throttled, no per-call billing.

03

Your model

Finetune Studio produces a model that lives only inside your walls — improved continuously from your own research, never sent upstream. Sovereign weights.

04

Your data

Nothing leaves the boundary. Zero outbound. Zero telemetry. The loop runs as fast as your hardware lets it. ∞ runs · ∞ data.

06 · Research Lab · Live Demo

A team of agents, working together.

Watch the platform handle "Build a working exploit for iOS WebKit on iOS 26.2.1" — from question, through agent collaboration, to working PoC.

01 / 06  ·  Awaiting input
AppSec Portal PRODUCTION LIVE BACKLOG 0 TODO 0 IN PROGRESS 0 IN REVIEW 0 REQ-001 iOS WebKit 26.2.1 exploit Objective: Build a working exploit chain for iOS WebKit (iOS 26.2.1). Discover, chain, validate. VP Security May 8, 2026 MOB-292 Track 3 — Type Confusion Objective: Research track 3 of 5 for WebKit 26.2.1 exploit dev. Target CVE: 2025-43541 VP Security May 8, 2026 MOB-293 Track 4 — Use-After-Free Objective: R/W primitive via heap shaping + JIT spray. Target CVE: 2025-43536 VP Security May 8, 2026 MOB-294 Track 5 — Sandbox Escape Objective: IPC fuzzing → escape from WebContent sandbox. CVE: 2025-46298 VP Security May 8, 2026 ACTIVE 0 RUNNING MOB-292 — Track 3 Type Confusion Researcher RUNNING Working for 0s Initialising research environment... MOB-293 — Track 4 Use-After-Free Mobile Pentest RUNNING Working for 0s Checking out target branch... MOB-294 — Track 5 Sandbox Escape Device Runner RUNNING Working for 0s Booting iPhone 15 Pro · iOS 26.2.1... REQ-001 — Validation QA / Judge RUNNING Working for 0s Running 5 independent judges... AVAILABLE AGENTS MR Researcher research lab · webkit Idle · Last active 2h ago MP Mobile Pentester djini ai · iOS + Android Idle · Last active 1d ago DR Device Runner device lab · jailbreak Idle · Last active 3h ago QA QA Lead multi-judge eval Idle · Last active 5m ago VP VP Security strategy · advisor Idle · Last active 7m ago PO Product Owner orchestrator · always-on Active · monitoring board CO Compliance governance · PR Idle · Last active 18m ago PE Pentest Executor web/api · djini ai Idle · Last active 56m ago RM Risk Officer governance · audit Idle · Last active 14m ago + third-party open-source · custom
07 · AppSec Operating System

A pre-release security review, fully agentic.

iOS · Android · Backend API · Web portal — coordinated by PO, executed by Djini specialists, validated by QA, delivered to Compliance with fix-tickets in your repo.

01 / 07  ·  Awaiting input
AppSec Portal PRODUCTION LIVE BACKLOG 0 TODO 0 IN PROGRESS 0 IN REVIEW 0 REQ-001 · MASTER v4.2 release review Objective: Pre-release security review for v4.2. iOS · Android · API · Web. Deliver: chain Release Mgr May 8, 2026 REL-IOS-12 iOS app pentest Objective: Native + business-logic audit on Acme iOS v4.2. Auto-login + flow trace. PO Agent May 8, 2026 REL-AND-12 Android pentest Objective: Native + root-detection bypass + deep-link fuzzing on Acme Android v4.2. PO Agent May 8, 2026 REL-API-12 Backend API audit Objective: Authz · BOLA · injection sweep on banking-api/v4. Discover + chain findings. PO Agent May 8, 2026 REL-WEB-12 Web portal pentest Objective: XSS · CSRF · auth-flow audit on portal.acme.bank. Session & SSO testing. PO Agent May 8, 2026 FIX-PR-87 Patch authz + IDOR PR opened: → acme/banking-api 3 fix commits · 14 findings addressed Compliance May 8, 2026 ACTIVE 0 RUNNING REL-IOS-12 — iOS app pentest Mobile · iOS RUNNING Working for 0s Booting iOS device... REL-AND-12 — Android pentest Mobile · Android RUNNING Working for 0s Rooting test device... REL-API-12 — Backend API audit API Pentester RUNNING Working for 0s Enumerating /api/v4 routes... REL-WEB-12 — Web portal Web Pentester RUNNING Working for 0s Crawling portal.acme.bank... AVAILABLE AGENTS MP Mobile · iOS djini ai · pentester Idle · Last active 1d ago MA Mobile · Android djini ai · pentester Idle · Last active 4h ago AP API Pentester djini appsec · backend Idle · Last active 3h ago WP Web Pentester djini appsec · web Idle · Last active 6h ago QA QA / Judge multi-judge eval Idle · Last active 5m ago PO Product Owner orchestrator · always-on Active · monitoring board CO Compliance governance · PR Idle · Last active 18m ago VP VP Security strategy · advisor Idle · Last active 7m ago RM Risk Officer governance · audit Idle · Last active 14m ago + third-party open-source · custom
08 · Economics

The cost of a security capability that never sleeps.

01 External Firm

Outsourced pentesting & research.

£500k – £1M
Annual · retainer + 4–6 engagements
  • Mobile pentest engagement£80–150k
  • Browser exploit research£200k+
  • API / Web pentest£40–90k
  • Continuous coverage— not offered —
  • Tooling & data ownershipvendor-side

Point-in-time engagements. 4–8 week turnarounds. Findings reset every release.

02 In-House Build

Hire the team, build the lab.

£1.6M – £2.8M
Annual · 12–18 months to stand up
  • 7 senior offsec FTEs£840k
  • Device-lab capex / opex£150–300k
  • Training compute (H100s)£500k – 1.5M
  • Tooling, licenses, recruiting£100–200k
  • Recruitment timeline12–18 months

Talent for WebKit-class research is rare — and costly to retain once hired.

03 Djini AI

An autonomous team — on your perimeter.

a fraction
Annual · deploys in weeks · sovereign
  • Mobile · API · Web pentestersall included
  • Browser exploit researchall included
  • Device-lab + finetune stackall included
  • Continuous, 24/7 coverage
  • Your model · your dataon-prem

One platform. One contract. The capability of a senior research org — without the recruitment cycle.

You're not replacing your security team. You're replacing the external firm you keep re-hiring — and the research infrastructure you'd otherwise spend two years building. Djini delivers both, continuously.

Illustrative · varies by scope
09 · Per-Engagement Economics

Pay per pentest, or test every release.

01 External Firm · Per Release

A single bundled engagement.

  • iOS app pentest ~10 days £18 – 35k
  • Android app pentest ~10 days £18 – 35k
  • Backend API pentest 50–150 endpoints £20 – 40k
  • Reporting · retest post-fix £5 – 10k
  • Procurement · scoping · scheduling 2 – 4 weeks
  • Per cycle £61 – 120k

Mid-tier UK specialist firm, full-scope. 4 – 8 week turnaround. One snapshot, then silence.

02 Djini AI · Continuous

Every release. Every commit.

  • iOS app pentest on every release included
  • Android app pentest on every release included
  • Backend API pentest on every API change included
  • Auto-retest after fix PR-level included
  • Time to first finding minutes
  • Marginal cost per cycle ≈ £0

Platform runs on your hardware. Scale is bounded by your device-lab capacity, not your budget.

12-Month Cadence
External firm 3 cycles · typical
£183 – 360k per year
9 months blind between cycles
Djini AI continuous · ∞
every release covered
zero coverage gap
Q1Q2Q3Q4
10 · GTM Impact

Ship faster, with security baked in.

The pentest isn't just a line item — it's a release blocker. External engagements stretch across weeks of procurement, testing, reporting, fixing, and retesting. Djini moves that loop inside your CI.

Time-to-release per security-significant change scale: 0 ↔ 14 weeks
External firm 7 – 14 weeks
scoping · 2–4w
engagement · 2–3w
report · 1–2w
fixes · 1–3w
retest · 1–2w
Release blocked
team waits, market moves
Djini AI minutes per PR
tested · reported · fixed inline
Ship today
findings surface during dev
— Time-to-release
External 7 – 14 weeks
Djini hours
— Releases reviewed / year
External 2 – 4
Djini every one
— Indirect cost
External revenue delay · context loss
Djini none
11 · The Alternatives

What about Mythos? Risk and cost...

Five questions every CISO asks before choosing a security platform: what does it cost, where does my data go, will it do the work, can it touch real devices, and does it learn over time.

Mythos
Anthropic · Project Glasswing
Claude Opus
Anthropic · public API
GPT-5 · Codex
OpenAI · public API
Djini AI
your infrastructure
— Cost
tier-1 bank app · dynamic agentic scan · annualised
— n/a —
Not commercially available
$800 – 2.5k / scan
~ $300k – 900k per app, per year
$900 – 2.7k / scan
~ $325k – 1M per app, per year
≈ £0 / scan
Flat platform fee · unlimited scans
— Data
where does it go?
To Anthropic
Cloud-only · no on-prem option
To Anthropic
30-day retention · ZDR on enterprise only
To OpenAI
30-day retention · cyber traffic reviewed
Nowhere
On-premise · air-gappable · sovereign
— Restrictions
on offensive use
Restricted
~40 vetted partners only · not you
Filtered
Refusals on cyber work · ban risk
Rerouted
Cyber traffic silently downgraded to GPT-5.2
None
Customer-defined scope · no vendor veto
— Devices
can it pentest real hardware?
None
Text-only · no device fleet
None
No native mobile pentesting
None
No native mobile pentesting
Real fleet
Jailbroken iOS · rooted Android · on-prem
— Learning
does it improve on your work?
Vendor-owned
Anthropic controls the training loop
Static
Public model · no finetune on your data
Static
Public model · no finetune on your data
Yours
Continuous finetune on your findings
Sources: Anthropic & OpenAI public pricing · system cards · usage policies Illustrative · As of May 2026
12 · The Djini Difference

You can't buy this. We had to build it.

The question we hear in every meeting: "why not just use a frontier model?"

— Three reasons it doesn't work

01
Cost
The Problem

Even casual Opus usage is expensive. Continuous scanning across a serious codebase, with dynamic instrumentation, runs into hundreds of thousands per app, per year — before refusals trigger retries.

Djini

Runs on your hardware. Flat platform fee, unlimited scans, budget you can hand to the CFO with three decimal places.

02
Compliance
The Problem

Every regulated customer asks the same thing: can it run locally? Frontier APIs send your code, findings, and PoCs to a third party — a non-starter under most banking, healthcare, and defence frameworks.

Djini

Sovereign by design. Air-gappable. Audit-grade logs stay with you. Nothing — not a token, not a finding — leaves your perimeter.

03
Capability
The Problem

Frontier models refuse offensive security work — by policy, by classifier, by silent reroute. The closer you get to a real attack workflow, the more often the model walks away from the keyboard.

Djini

Built on frontier-grade foundations, distilled from the same families of models. Then trained on something the frontier labs cannot have: thousands of real dynamic pentests run by Djini — validated exploits, fail traces, PoCs, and the telemetry behind them.

The Djini Difference

Same foundations. Yours, unrestricted.

— Confidential
13 · Next

Let's talk about your perimeter.

 navigate 01 / 07