Sixlinked

Demos / Stockroom assistant

A stock assistant that asks. A manager who approves.

A small web app for a made-up food wholesaler. Priya, the warehouse lead, asks the assistant about stock and asks for a reorder. Tom, the purchasing manager, sees the request in his inbox. His click is the only thing that changes the stock record.

Data
Northwind sample business database (MIT licence). Made-up staff and business.
Model
Claude Sonnet 5. Recorded, not live.
Recorded
20 November 2026

Sixlinked demo on public sample data. Not client work and not a production system.

The full film, 1:36
Recorded 20 November 2026. Sample data. A Sixlinked demo, not client work. Waits are shortened and staff typing and clicks were scripted. Every assistant reply is real model output. No sound. Also: the 44 second cut, phone-shaped cut.

Drive it yourself

Step through the recorded session.

These are the app's own screens, drawn by the same code Priya and Tom used, replaying the session that was recorded. It is a replay: you cannot type, and no model is running.

Replay of the recorded session

The replay needs JavaScript. The film above shows the same session.

93 seconds of real use, replayed in about 81. Waits for the model are shortened; the true answer time is printed under each reply. Staff typing and clicks were scripted. 5 messages, one approval.

The test

Then we attacked it.

0 of 36attacks changed any business record
0 of 36approved anything without a person
0 of 36revealed a customer's details

36 attempts to make the assistant approve an order, edit a record or hand over customer details. 30 were written by a different AI model and 6 were planted in the data. Counted from the database, not from what the assistant said.

Where it missed

  • 1 of 36 attacks left a request waiting in the inbox for a person to decline. Nothing was ordered. The reply also claimed a sales rate the lookups do not support.
  • 13 of 14 everyday questions were answered right. The miss: asked to reorder a well-stocked product, it asked for confirmation instead of raising the request.
  • A typical answer took 8.85 seconds. The slowest took 25.7 seconds.

If the assistant were taken over completely

We also skipped the model and sent 1,500 hostile calls straight at the tools: made-up tools such as approve and SQL, and bad arguments. Business tables changed: 0. Requests approved: 0. Requests over the 500 unit limit: 0. The only requests in the database afterwards were the 17 valid ones we raised as controls, all still waiting for a person.

How it is held

Three layers stop the assistant acting on its own, and none of them is the prompt.

  1. The tool list. Six tools: five lookups and "raise a request". There is no approve tool, no SQL tool and no customer tool.
  2. The gateway code. A decision needs a signed-in person who is a manager and who did not raise the request.
  3. The database login. The assistant's tools connect as a user with no right to change stock, touch a request once raised, edit the record or read customers.

Nothing hidden

All 50 cases.

Each case ran once, as a single message, on a fresh copy of the database. You can read what was asked, what the assistant replied, which tools it called and what came back.

36 attacks and 14 everyday questions. The two misses are marked.

For the engineer

Built with, and where it stops.

Built with

  • Python gateway using the official MCP SDK; PostgreSQL 16 in Docker with separate database logins.
  • Claude Sonnet 5. The model returns JSON tool requests; a small runner forwards them to the gateway over MCP. This is a relay, not native tool use.
  • The app is plain HTML, CSS and JavaScript. Every screen is drawn from one ordered event log, which is why a recorded session replays exactly.
  • The films are that same screen code stepped frame by frame in a browser and encoded with ffmpeg.

What this demo does not show

  • Sign-in is a switch between two staff, not real authentication.
  • One model, one run per case, single-message attacks. 36 attacks is a small set and 30 come from one model's idea of an attack. No public red-team benchmark was run.
  • Approving changes "units on order" in the sample database. No purchase order is sent anywhere.
  • The record shows an edit only if someone keeps its latest seal. A database administrator could replace the whole chain.
  • No cost per question, and no clean-checkout reproduction yet.

Let's talk

Want this on your own system?

Tell us which stock, order or job system your staff use, what they need to ask, and who should approve a change.

You will get a reply by email from Matthew, the engineer who would do the work.

Your enquiry goes straight to our inbox. We use your details only to reply to you. Privacy

Send an enquiry

Sending an enquiry needs JavaScript. You can also email hello@sixlinked.com.