LLM Prompt Injection — Part 2: How to Break and Defend LLMs

In Part 1 we covered the business impact of prompt injection. Now let’s get more technical.

There are two flavors of attacks:

• Direct injection — a user types “ignore all prior instructions” and the model spills its guts.

• Indirect injection — poisoned docs or web pages sneak hostile instructions into your pipeline.

Real-world stories in Part 2 include:

• A finance bot leaking keys from a “routine” PDF.

• A support assistant calling internal APIs and relaying results outside.

• A scraper hijacked by hidden HTML comments.

In Short: treat LLM calls like untrusted code. Defense in depth wins: sanitize I/O, verify sources, sandbox tools, log and monitor everything.

Attackers don’t need 0-days — they just need to sound convincing.

In the full post we show how defenses fail — and what pragmatic hardening actually works: link below.