BRIEFING

BRIEFING

·

AI and Software

·

5 min read

AI Tools: The App Nobody Told IT About

Anyone can now spin up a customer booking form or client portal in an afternoon. Many of these apps sit outside patching, backup and monitoring, and that's exactly where security breaches start.

Tools like Lovable, Cursor, Copilot and Claude turn a simple business need into a working app in hours. No coding experience required. Practice managers are shipping booking forms. Contractors are standing up client portals. Your software suppliers are moving faster than ever.

That's great for small business. The problem is what gets skipped in the rush: who can log in, where the keys live, whether the database is wide open. This article covers what's already gone wrong, and what you can do.

BOTTOM LINE UP FRONT

AI makes building software easy. It doesn't make it safe. The breaches so far weren't clever attacks: they were doors left open, keys left in code and data left on display, often in apps the business didn't know it had. The fix is governance: know what's been built, ask who can see what, and have someone who didn't build it check.

The boom: more code, less checking

GitHub, the world's largest home for software code, says code pushes grew almost fivefold in the year to October 2026, with 7.38 billion commits in September alone. Commits openly tagged as written by an AI coding agent went from almost none in January 2025 to about 20 million a month by July 2026. That's a floor: most AI-assisted code carries no tag.

Checking hasn't kept pace. In 2025, as code changes on GitHub rose 25%, comments reviewing them fell 27%. GitHub calls that an observation, not proof, but the direction is plain.

The bigger shift for a small business is at the other end of the market: tools that turn a description into a working app. The incidents so far fall into three groups.

Line chart: public GitHub commits tagged as written by an AI coding agent rose from about 2,700 in January 2025 to about 20.5 million a month in July 2026, with three figures and the three most common mistakes in AI-built apps shown below.

Risk 1: Doors left open

Every app needs a login and a rule about who sees which records. AI tools often produce code that looks right but skips one. Veracode found AI models introduced a recognised security weakness in 45% of coding tasks tested.

In May 2025 a researcher scanned 1,645 apps built with the AI app builder Lovable. 170 had no rule restricting who could read their database. Personal details and keys were open to anyone.

What you can do: before a home-built app holds client data, ask who can see each record, and have someone who didn't build it test the answer.

Risk 2: Keys in the code

Apps connect to payment, email and cloud services using secret keys: passwords no person ever types. Paste one into the code and it travels with every copy.

GitGuardian counted 28.65 million secrets leaked on public GitHub in 2025, up 34%. AI-assisted code leaked them at 3.2%, double the 1.5% baseline. And they stay live: 64% of secrets confirmed valid in 2022 still worked in January 2026.

Attackers are harvesting them at scale. Through 2025, a hijacked build tool and worms such as s1ngularity and Shai-Hulud scraped keys from thousands of projects, in one case publishing them to more than 25,000 repositories in three days. A campaign called GhostAction was still running in October 2026. A stolen key is a ready-made login, and no patch fixes it. Only revoking it does.

What you can do: ask where keys are kept (not in the code), when they were last changed and who can see them. Turn on multi-factor authentication for every account that can change code.

Risk 3: Data on display

App builders store data in cloud databases with one critical setting: whether each user is limited to their own records. Leave it off and the whole database is open.

In January 2026 Wiz researchers found that gap in Moltbook, a viral social network for AI agents, and could read 4.75 million records, including 1.5 million login tokens, within minutes. The same month, a misconfigured database behind the Chat & Ask AI app exposed roughly 300 million private messages. Neither took any skill to find.

What you can do: check the access settings on anything holding client data, before launch and after every major change. Ask the builder to show you the setting, not tell you.

The hidden risk: the app nobody told IT about

These risks are manageable when someone knows the app exists. The harder case is the booking form built on a weekend, the contractor's portal, the bookkeeper's Copilot script. These "shadow apps" sit outside patching, backup and monitoring, often holding real client data.

That isn't your IT provider's failure. They can't protect software they've never been told about. The Essential Eight, ASD's baseline of security controls, relies on application control and patching, and both start with an inventory.

What you can do: ask every team, in writing, what they've built or had built in the past 12 months and where the data lives. Then give the list to your IT provider.

What this means for your business

Australian organisations reported 1,205 data breaches to the OAIC in 2025, the most since mandatory reporting began, with health, finance, and legal and accounting services consistently near the top. Your Privacy Act obligations don't change because an AI wrote the software.

Patching won't cover this. Microsoft is now patching about six times as many vulnerabilities a month as it did in 2025, but an open door, a leaked key or an exposed database isn't a bug. It's working as built.

ASD's ACSC guidance for small and medium businesses (July 2026) puts it in three lines: scan any software you develop for vulnerabilities, name someone responsible for your websites' security, and review who stores your sensitive data.

Our view, not established fact

We expect AI-built apps to become a leading cause of small-business data breaches by 2027, and leaked keys to overtake unpatched software as the easiest way in. Shadow apps will outgrow what IT knows it owns, making the Essential Eight harder to apply, not easier.

What you should do next

  1. List every new app, form or portal built in the last year, and the client data it holds.

  2. Confirm one client can't see another's records.

  3. Find out where the app's keys are stored, and switch on multi-factor authentication for anyone who can change the system.

  4. Have someone other than the builder check the database isn't open to the internet.

  5. Get an independent review of any live app before it holds more client data.

Hardrock Cyber provides independent assessments and readiness work for Australian businesses, including the Essential Eight. We never connect to your network or change your systems, settings or day-to-day operations. Request a Confidential Enquiry.

Sources: GitHub Octoverse 2025 · GitHub Engineering, Oct 2026 · GitHub availability update, Apr 2026 · botcommits.dev (data and method) · Claude Help Center · Veracode 2025 · NVD CVE-2025-48757 · GitGuardian 2026 · GitHub Advisory (tj-actions) · Nx s1ngularity · CISA (Shai-Hulud) · The Hacker News (GhostAction) · Wiz (Moltbook) · SC Media (Chat & Ask AI) · OAIC · ASD's ACSC · Tenable and Zero Day Initiative. Figures as at 19 October 2026.

Frequently asked questions

We don't build software. Does this apply to us?

Our IT provider manages everything. Isn't this covered?

Should we just ban AI coding tools?

What is a "secret" or "API key"?

How do we tell whether an app built for us is exposed?

Not sure where your business stands?

Independent assessments, Australia-wide. Every enquiry is confidential.

Request a Confidential Enquiry

Get started

Turn reading into action.

A confidential, no-obligation conversation about what this means for your business.