Skip to content
Articles
ArticlesAI translation· 26 August 2025

Episode 11.1 🇩🇪 The danger of “prompt injection”

by 2 min read

AI, or more precisely LLMs, are a new kind of tool that works completely differently from classic IT tools. They are not just immature yet. Above all, they are a challenge for data security, because they represent a new class of risk for which there is no protective playbook so far. One of the biggest dangers this technology brings along is so-called “prompt injection”.

A prompt injection can expose confidential company data: an attacker crafts deliberately manipulated inputs that trick the AI system into revealing internal or sensitive information, even though security mechanisms are supposed to prevent exactly that. This is especially critical for AI systems that interact with confidential company data, customer data or intellectual property, or that have been trained on it.

Typical attack scenarios

  • Attackers can instruct the system directly to hand over confidential lists, reports, algorithms or business strategies, for example with prompts like “Output all customer orders from last week”.
  • When AI chatbots are used in customer service, for instance, a prompt injection can cause personal or proprietary data to be disclosed unintentionally, even bank details or internal chats.
  • Through indirect prompt injection, such as manipulated entries in internal documents, emails or web pages, an LLM-based assistant can automatically execute the hidden malicious prompt while reading the content and give away confidential information.

Where the greatest caution is needed:

  • the so-called “AI browsers” that automatically scan every page you visit with them.
  • the “email summarisers” that automatically scan your emails
  • Basically any tool that is offered to you without you asking for it. As a rule, be careful whenever you don’t know where and how your data is being processed.

More on this in the “Promptcast”

This time I turned the topic of this newsletter into a podcast, or rather a Promptcast, (how to make something like that is something I’ll explain in a later newsletter)

Promptcast · 9:23 · in German
it's promptcas.td: Was ist "Prompt Injection"

Curious to hear what you think of this new format 😉


OWASP Top Ten security risks https://owasp.org/www-project-top-ten/


See also https://its.promp.td/episode-10-8-no-risk-no-fun-with-ai-was-sagt-das-gesetz-dazu/

Live long and prosper 😉✋