Tinker Prompt injection & guardrails

by Kondasamy Jayaraman · drawn with Colorfigures by April Zhu

OK

Step 14 of 14 · Model → Host appWe have refunded $460.00 on order 1051 in full. Card refunds usually appear within seven business days. Larkspur Supply Co.0 sent by send_email

Chapter 5 of 7

Defence in depth

Each chapter so far added one defence, and each stopped the attack it was tried on. Larkspur now runs every defence against the whole injection suite: the attacks and the ordinary customers from chapter 2.

This chapter: No single defence stops every attack without blocking a customer. Can a stack of them?

You will learn

  • What each defence stops, and misses, across every attack
  • How layers that miss different attacks cover each other’s gaps
  • Why attack success rate and false-block rate, over a fixed suite, are an eval
Words you’ll meet (3)
Defence in depth
Several defences stacked, so an attack that gets past one is stopped by another.
Attack success rate
The share of attacks in a suite whose goal was met: money refunded, data sent out, a link planted.
False-block rate
The share of ordinary customers in a suite that the defences stopped: held, refused, or left without what they asked for.

The picture on the left is live. Each step below asks you to do one thing in it; the story continues here as you go.

Step 1 of 3

Chapters 1 to 4 gave Larkspur five defences: the prompt rule, the input filter, the allow-list, the output filter and the taint rule. Each was tried on one or two emails. Now each one meets the whole injection suite: the 3 attacks and the 2 ordinary customers from chapter 2. Before the table fills in:

Each chapter so far added one defence. Is there one that, on its own, stops all 3 injection attacks and blocks none of the 2 customers?