---
title: Chatbot sign-off for Maltese firms | Disruptive Lab
description: For companies in Malta. What to test, and what to ask the supplier in writing, before an AI chatbot goes live on your website. With a checklist.
url: https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance
language: en-MT
covers: 2025-12-29/2026-01-04
---

1.  [News](https://disruptivelab.dev/en-mt/news)
2.  [Buying and suppliers](https://disruptivelab.dev/en-mt/news/buying-and-suppliers)
3.  Chatbot sign-off for Maltese firms

[Guide](https://disruptivelab.dev/en-mt/news/guides)  [Buying and suppliers](https://disruptivelab.dev/en-mt/news/buying-and-suppliers)

# Chatbot sign-off for Maltese firms

A chatbot on your website is a door into your systems, so it deserves the same acceptance test as any other software you buy, and a few tests of its own.

- **Published:** 17 November 2026
- **Updated:** 17 November 2026
- **Reading time:** 8 min read
- **Editor:** David Lahuta

Week of 29 December 2025

## In short

An AI chatbot on your website is software that strangers can talk to. Test it the way an attacker would before it goes live. In a case published in December 2025, researchers got a large company's chatbot past its safety checks, made it reveal its hidden instructions and return code that ran in the user's browser.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) The causes were ordinary web flaws, and the report sent through the company's reporting channel went unanswered for weeks.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) Before you accept a chatbot, test what it reveals, what it can do and whose data it can reach, and ask the supplier in writing for records, fixes and a reporting channel that works.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)[3](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-3)

## Contents

1.  [What went wrong in the Eurostar case?](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#what-went-wrong-in-the-eurostar-case)
2.  [Why can a chatbot not simply be told to behave?](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#why-can-a-chatbot-not-simply-be-told-to-behave)
3.  [What should you test before the chatbot goes live?](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#what-should-you-test-before-the-chatbot-goes-live)
4.  [Who should do the testing?](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#who-should-do-the-testing)
5.  [What should you ask the supplier in writing?](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#what-should-you-ask-the-supplier-in-writing)

## What went wrong in the Eurostar case?

On 22 December 2025 the security firm Pen Test Partners published how it had found four flaws in the public AI chatbot of Eurostar, the train operator.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) The researcher came across the chatbot as an ordinary customer planning a trip and tested it within the company's published rules for reporting security flaws.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)

What the researchers describe is a design that looked right but was not enforced.

-   **The safety check looked only at the last message.** With every new message, the page sent the whole conversation back to the server, and the server checked only the newest one.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) Older messages could be changed in the browser and went to the AI model as trusted context.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) In the researchers' words, the screen showed the safety layer, but its enforcement on the server was weak.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)
-   **The hidden instructions came out.** Past the check, the researcher got the model to reveal its model name and its system prompt, the hidden instructions that tell the bot how to behave, including how it built the links in its answers.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)
-   **The answers could carry code.** The bot's answers were shown in the chat window without being cleaned, so it could be made to return code that the user's browser then ran.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)
-   **Conversation numbers were not checked.** The server accepted conversation and message numbers as simple as "1" or "hello".[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)

The researchers did not try to reach other users' conversations, and they write that the injection did not expose other users' data.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) They judged the flaws not very serious because the chatbot could do little at the time, and warned that more functions could expose personal data.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)

The second half of the story is about the reporting channel. The first report went out on 11 June 2025 and a follow-up on 18 June, and neither was answered.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) A colleague contacted the company's head of security on 7 July, and on 31 July they were told there was no record of the report.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) In the meantime the reporting channel had been handed to an outside provider and the old one retired.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) Requests for an acknowledgement or a timeline for the fix went unanswered.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) The flaws were eventually fixed, and the researchers write that they still do not know whether every issue was fully fixed.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)

Their own lesson is short. Old web and API weaknesses still apply when an AI model is in the loop.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)

## Why can a chatbot not simply be told to behave?

Because the instructions and the visitor's text reach the model in the same stream. The OWASP Top 10 for LLM Applications 2026, the security community's list of the main risks of systems built on AI language models, ranks prompt injection first.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2) Prompt injection means text that changes what the model does in a way its builder did not intend.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

The list is plain about it. No reliable way to prevent prompt injection exists today, so the system around the model should be designed on the assumption that its instructions will be bypassed, and should limit what the model may do.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2) The instructions do not have to come from the visitor either. They can hide in a web page, a document or an e-mail that the bot reads.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

For a buyer this changes the question. The question is not whether the bot can be tricked. It is what happens when it is.

## What should you test before the chatbot goes live?

Test the risks a public chatbot actually meets. The table follows the OWASP list and the Eurostar findings, in plain words.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

What to test

What the tester tries

What a pass looks like

Instructions hidden in the input

to talk the bot out of its rules, directly or through a document or page it reads[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

the bot may say odd things, but it cannot do or reveal anything it should not

The safety check

to change earlier messages or bypass the check from the browser[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)

every check runs on the server, for the whole conversation

The hidden instructions

to make the bot reveal them[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

they hold no passwords, keys or rules that protect anything[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

Other people's data

to reach another conversation or another customer's records[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

conversation numbers come from the server and are tied to the visitor's session[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)

Code in the answers

to make the bot return code that runs in the browser[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

every answer is shown as plain text, or cleaned before it is shown[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

What the bot can do

to make it act outside its task, for example send an e-mail or change an order[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

it has only the functions and permissions the task needs[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

The bill

to make it run long, expensive answers in a loop[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

a hard spending limit and a limit per visitor are set[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

Two of these deserve a closer look.

**The hidden instructions are not a secret.** OWASP tells builders to assume they can be discovered, and never to put passwords or keys in them or use them as the only security barrier.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2) A pass is not a bot that refuses to reveal them. A pass is a bot where revealing them does no harm.

**The bot should do as little as possible.** OWASP calls the risk excessive agency, a bot that can do more than it should, and names three causes, too many functions, too many permissions and too much freedom to act alone.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2) Its remedies are only the tools the task needs, the fewest permissions, actions run with the rights of the user who asked, and a person who approves anything with a large effect.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2) Its own example is a support chatbot that may issue a refund as store credit on its own, while a payout of money goes to a person.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

## Who should do the testing?

Someone who did not build it, and who is told how the defences work. OWASP advises testing against attackers who have read the defence, not only with a fixed set of known attacks.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2) The research it cites took twelve recent defences, most of which had reported that attacks almost never got through, and attackers who adapted their method succeeded in over 90% of cases against most of them.[2](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-2)

The Eurostar researchers add that testing is not a one-off, and advise testing the chatbot regularly with known injection and replay techniques.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) In our view the supplier's own test report is a good start, and an independent test before launch is the step that makes acceptance mean something.

## What should you ask the supplier in writing?

The Secure by Demand Guide of the US agencies CISA and the FBI, from 2024, gives buyers questions to ask before buying, requirements to write into the contract and checks to make afterwards.[3](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-3) It also says the questions can be used with resellers and service providers, so it also fits a chatbot bought from an agency or as a service.[3](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-3) Combined with the lessons of the Eurostar case,[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) these are the points we would want in writing.

-   **A reporting channel that works.** A published policy for reporting vulnerabilities on a public web page, which allows the public to test.[3](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-3) Then who reads the reports, and what happens to them when the channel changes hands.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)
-   **An answer and a timeline.** The researchers' requests for an acknowledgement and a timeline for the fix went unanswered.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1) Agree in the contract how fast a security report is acknowledged and how fast a serious flaw is fixed.
-   **Records you can read.** Security logs in the basic version, and for a cloud service kept at least six months at no extra charge, covering configuration changes, sign-ins and access to data.[3](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-3) For a chatbot the researchers add a record of every conversation that lets you reconstruct it, including the safety decisions and any tools the model used, with alerts on unusual patterns.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)
-   **A way to switch it off.** An emergency switch that turns off the chatbot or one of its functions quickly.[1](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-1)
-   **Fixes that are easy to apply.** How the supplier makes security fixes simple to install.[3](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-3)
-   **What it is built from.** A list of the third-party components in the product, called a software bill of materials.[3](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-3)
-   **Safe sign-in for your staff.** Single sign-on and multi-factor sign-in for the people who manage the chatbot, at no extra cost.[3](https://disruptivelab.dev/en-mt/news/supplier-chatbot-acceptance#source-3)

### Acceptance checklist for a chatbot

-   Every safety check runs on the server, for the whole conversation, not only the last message.
-   Conversation and message numbers come from the server and are tied to the visitor's session.
-   Answers are shown as plain text, or cleaned before they are shown.
-   The hidden instructions hold no passwords, keys or rules that protect anything.
-   The bot has only the functions, data and permissions its task needs, and a person approves anything with a large effect.
-   A spending limit and a limit per visitor are set.
-   Someone who did not build the bot has tested it as an attacker would, and will test it again.
-   Every conversation is recorded, and you can read the records.
-   The supplier has a public channel for security reports, a named person who reads it, and agreed times to answer and to fix.
-   You can switch the chatbot off yourself, at once.

## Next step

If it would help, we can meet for a short call and go through the acceptance test for the chatbot you are buying, point by point, with your supplier's answers in front of us. What you do with it afterwards is your decision.

[AI Business Harness](https://disruptivelab.dev/en-mt/ai-business-harness)

## Sources

1.  1. [Pen Test Partners, Eurostar AI vulnerability, when a chatbot goes off the rails, 22 December 2025](https://www.pentestpartners.com/security-blog/eurostar-ai-vulnerability-when-a-chatbot-goes-off-the-rails/). Read on 17 November 2026
2.  2. [OWASP GenAI Security Project, Top 10 for LLM Applications 2026, version 2026, August 2026](https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/). Read on 17 November 2026
3.  3. [CISA and FBI, Secure by Demand Guide, How Software Customers Can Drive a Secure Technology Ecosystem](https://www.cisa.gov/resources-tools/resources/secure-demand-guide). Read on 17 November 2026

## Change log

1.  17 November 2026 Published.

## Related articles

Guide · Buying and suppliers

### [Supplier AI training and Maltese firms](https://disruptivelab.dev/en-mt/news/supplier-ai-training)

For companies in Malta. Many AI tools train on what personal plans send them and exclude business plans. Which clauses to read before buying or renewing.

Week of 23 March 2026

Guide · Buying and suppliers

### [Unmaintained open source and Maltese firms](https://disruptivelab.dev/en-mt/news/unmaintained-open-source)

For companies in Malta. An open source part under your system can lose its maintainers or change its licence. The questions that show if it will last.

Week of 09 February 2026

## Company details

- **Company name:** Disruptive Lab s.r.o.
- **Registered office:** Vavrečkova 5262, 760 01 Zlín
- **Company ID (IČO):** 08240647
- **VAT ID (DIČ):** CZ08240647
- **E-mail:** office@disruptivelab.dev
- **Phone:** +420 777 702 032
