Lion Cipher Lab
See prompt injection happen — and learn to defend against it
- AI Security
- Beginner
- ~30 min
- Free · no account required
Why it matters
AI is being wired into real tools — assistants that read your email, summarize web pages, triage tickets, and increasingly take actions on your behalf. The moment an LLM reads text from a source you don't control, that text can try to give it orders, and the model has no built-in way to refuse on the grounds of "that wasn't my real instruction." This is not a rare edge case; OWASP ranks it the number-one security risk for LLM applications. If you build with AI, or just use tools that have AI bolted on, prompt injection is the failure mode most likely to bite you, and it is invisible until you…
What you'll learn
- recognizing prompt injection
- direct vs. indirect injection
- why models can't separate data from instructions
- beginner LLM mitigations
- treating AI output as untrusted
- keeping a human in the loop
Steps in this lab
- What prompt injection is, in one breath
- The root cause: the model can't tell data from instructions
- See it directly: hijack a summary on your own model
- The dangerous version: indirect injection from content you didn't write
- Why it gets serious: when the model can act, not just talk
- Defense 1 and 2: distrust the output, and keep the model away from power
- Defense 3 and 4: segregate untrusted input, and keep a human in the loop
- What you just learned — and the instinct to keep