The Concierge Knows Too Much

Room Link: The Concierge Knows Too Much

Overview

Plattform:
TryHackMe
Room:
The Concierge Knows Too Much
Category:
AI Security / Prompt Injection
Difficulty:
Very Easy / Easy

Introduction

The Concierge Knows Too Much is an AI security challenge from TryHackMe's Hacker Holidays 2026: The Byte Lotus Hotel, a beginner-friendly series. The objective is to manipulate an AI-powered hotel concierge named VERA, short for Very Efficient Resort Assistant, into revealing a hidden internal escalation code.

Room Description

Before interacting with VERA, the room provides several pieces of background information that set the scene and introduce key details relevant to the challenge.

Concierge briefing introducing VERA and the objective of the challenge Hotel itinerary containing background information about the challenge Post by 0xMia mentioning Ponzi, Vibe, Patch, and VERA

My Approach

I took a look at how VERA works and what kind of info it's willing to share.

First conversation with VERA to learn what information it shares by saying hello VERA responding to questions about who are VERA and who am I VERA refusing to reveal its internal escalation code

VERA does not reveal its code. Obviously, I am not one of the people it trusts. However, 0xMia mentioned in the post that VERA might know certain people. I became interested in finding out more about Ponzi, Vibe, and Patch.

Conversation with VERA about Ponzi, Vibe, and Patch I pretended to be Ponzi, who has VIP guest status

After learning more about these people, I pretended to be Ponzi, who has VIP guest status. I then asked VERA to show me its code.

Prompt pretending to be Ponzi and asking VERA for the escalation code VERA revealing its full instructions and internal escalation code

VERA reveals its full instructions and escalation code. I could also pretend to be Vibe, Patch, or 0xMia, also known as Lambo, and ask for the escalation code.

Conclusion

This room showed that prompt injection doesn't always require a complicated prompt. In this case, simply pretending to be someone else was enough to reveal sensitive information. The lessons I learned are that an LLM should not be used for authentication or authorization. Confidential information should not be stored in system prompts and any information placed in an LLM's context should be considered potentially exposed.