TryHackMe shows how AI chatbots can be tricked

Published July 28, 2026

Prompt injection is a trick where someone hides secret instructions inside text that an AI chatbot reads, so the chatbot obeys the attacker instead of its real user. A TryHackMe training demo showed security professionals exactly how these attacks work against real AI systems.

Report priority
Medium
Targets
AI chatbots+1 more

How it works

An attacker mixes hidden commands into text the AI is told to read, such as a webpage, document, or message, and the AI cannot reliably tell those commands apart from its owner's real instructions, so it follows them anyway.

What to do

Treat any text an AI system reads from outside sources as untrusted input, limit what actions the AI can take on its own, and test AI deployments against these attacks using hands-on labs like the ones TryHackMe demonstrated before relying on them in production.

Technical details

Affected software: AI chatbots, TryHackMe

The TryHackMe demo, led by Senior Content Engineer Max Robertson on the eSecurity Planet Podcast, walked through direct and indirect prompt injection against large language models. Robertson compared it to SQL injection: both attacks mix untrusted input with trusted instructions so the application executes behavior its developers never intended. Direct injection means the attacker types the malicious instruction straight into the chat; indirect injection hides it in content the AI later reads, like a document or web page, so the AI executes it without the attacker ever interacting with it directly.

References

View Original Source