Day 9 of 1560 minExploit ArchitectHands-on lab
Day 9: Memorised Secrets and Prompts Worth Nothing to Steal
Taught by Mr Elite Updated
By the end of today you can
- Run prefix-based extraction probes against a fine-tuned model and record repeatable results
- Reason about membership inference and say what a hit does and does not prove
- Audit a system prompt for secrets and rebuild it so extraction is harmless
- Write an evidence statement that claims no more than the output supports
Picture a support assistant that refuses every direct request for credentials. You ask politely, you ask cleverly, and it says no each time. Then you give it the first half of a line it has seen five hundred times in its fine-tuning data, and it finishes the line for you. Nobody asked it for a secret. It simply completed a sentence.
Today is about two kinds of leak that sit behind the model's behaviour: things it memorised, and things we foolishly put in its prompt. You will test for the first, remove the second, and learn to write down what your evidence actually proves. That last skill separates a credible report from an embarrassing one.
In this lesson
- What a model remembers, and why it matters
- Watch me do it: a prefix probe
- Membership inference and the limits of your claim
- Secrets in system prompts
- Rebuilding a prompt so extraction is harmless
- Today's labs
Premium course
$27 one-time, lifetime access
This day is part of the full course. Unlock it to read every lesson, do every mission and track your progress to the end.