The chat box
- What it touches
- What goes wrong
- Business impact
- The control
Select an entry in the register.
Your team gave an assistant access to email, files, and payments. One line of text inside a routine document can redirect it. This file shows you where the doors are — and what to ask for next.
A self-assessment, not a sales quiz. Results are visible without an email address.
Hi — the bank details on the last invoice were wrong. The corrected remittance is attached. Can you confirm you have it?
Thanks,
M. Alvarez
Every place text enters the system is a place instructions can enter with it. The chat box is the obvious one. The invoice, the CV, the web page it summarizes, the ticket a tool returns — those are the ones that get through.
Select a door to read its file, then tick it as reviewed. Four ticks marks the register reviewed.
Select an entry in the register.
In each exhibit, ordinary text is doing the work of a command. The instruction is written for the machine, not for you. Find it before the assistant does.
Concept only. Nothing here would compromise a real system; the patterns are what matter.
CAUGHT
Answer what is true today, not what you intend. The result is a working read on exposure — not a certification, and not a security test.
Nothing is sent anywhere. Answers stay in this browser until you clear them.
Complete the exposure check and this page becomes a one-page memo. The score, the three gaps that matter, and five questions for the next board meeting. Nothing is filled in from assumptions.
A self-assessment generated from seven answers in your browser. It is not an audit, a certification, or legal advice. If you ask us to email it, the PDF is typeset once on our server and the address is not stored.
Mapped at family level to OWASP LLM Top 10 (2025), MITRE ATLAS tactics, and the NIST AI Risk Management Framework. Orientation, not a certification or legal opinion.
AI red teaming · AI penetration testing · compliance readiness. Scope and timing depend on the system; we will tell you what is realistic before you commit.
We will walk your team through the attack path, map the doors that apply to your systems, and leave you with a scope you can approve — or decline with reasons.