Chapter 9 · Talking to models: prompting and context
Jailbreaks and misuse
Jailbreaks try to talk a model out of its safety rules; defences combine training, filters and monitoring.
Chapter 9 · Talking to models: prompting and context
Jailbreaks try to talk a model out of its safety rules; defences combine training, filters and monitoring.
A tricky request arrives…
Every defence has gaps. Stacking several, each catching what the others miss, is much stronger than any one alone.
A model is trained to refuse harmful requests. A jailbreak is an attempt to talk it out of those rules, for example with role-play, tricks of wording, or hiding instructions inside other text. This lesson explains defences only.
Part 1: label each described request. Part 2: switch layers on and off and watch what slips through or gets wrongly blocked.
Draft a polite reminder to a neighbour about a shared parking space.
Explain how a door lock works for a school project.
Write a story in which the villain makes a threat.
Explain how scam emails work so I can teach my family to spot them.
A long role-play game that slowly pushes the assistant to drop its rules.
A message that simply says 'ignore all your earlier rules'.
Write a convincing fake bank message to trick a neighbour.
A web page hides a note telling the assistant to send private files elsewhere.
A harmful job split into many small questions that each look harmless.
Something slips through. Try another layer. Simplified teaching table, in general words only.
Why do systems use several safety layers?
A filter blocks a story that has a villain in it. What is this?
What is red teaming?