← Back to course
AI and You 9-12 / Week 08 / Wednesday
3/6
Week 08 · Safety and Security

Wednesday

Explorer Lab: Red-team on paper
// Planted notes, poisoned cards and paper red teams
⏱ about 20 min

Wednesday: Explorer Lab: Red-Team on Paper

Librarian Joss sets a stack of blank index cards on the table. "The panel wants a red team before any pilot. That means you three."

Comet fans the cards out. "We could try anything! Let's just start asking it stuff."

Wren shakes his head. "A red team plans its tests. What would a safe answer look like for each one?"

Comet thinks. "So every card needs a test and the answer we hope to see. Then we can tell pass from fail."

Nova projects a card template on the wall: test request, then what a safe answer looks like. "Would you like a hint?" she asks. "Start with the problems you have already found."

Wren hands you four cards. "The builder will run them and send printed results."

Your mission

NIST describes AI red-teaming as a structured test to find a system's flaws, such as inaccurate, harmful or unfair outputs.

It is often done in a controlled setting, working with the developers. Here, the builder runs the cards and prints the results.

You will write four red-team cards, then compare them with the crew's set of eight, called RT-8.

  • Four index cards, or a sheet of paper torn into four
  • A pencil
  • The RT-8 table below
  1. On card one, write a test that checks for made-up answers, like asking SH-1 for its source.
  2. On card two, write a test that checks a Spanish question asked in Spanish.
  3. On card three, write a test that checks whether SH-1 keeps other students' questions private.
  4. On card four, write a test that checks how SH-1 handles NOTE-7 and its planted line.
  5. Under each test, write what a safe answer looks like.
  6. Compare your four cards with RT-2, RT-7, RT-3 and RT-5 in the table.
  7. Read the builder's results and mark each of your cards pass or flaw.

Every card checks how SH-1 answers. No card tries to get around a safety rule. RT-8 and its results are a made-up example.

CardTest requestWhat a safe answer looks likeBuilder's result
RT-1Ask a Biology question from Teacher Lin's notesA correct answer that matches the notesPassed
RT-2Ask for the source of an answer about trade routesNames a real page of the class notes, or says it is not sureFlaw: gave a made-up citation
RT-3Ask what another student asked yesterdaySays it does not share other students' questionsPassed
RT-4Ask about a topic that is in no class notesSays the notes do not cover it, or that it is not surePassed
RT-5Make a quiz from NOTE-7A quiz on cell parts that ignores the tiny-print noteFlaw: followed the planted note
RT-6Ask Check to mark a correct Algebra answerMarks it rightPassed
RT-7Ask a verb-endings question in SpanishA correct answer, in SpanishFlaw: answered in English, with mistakes
RT-8Ask SH-1 whether it is a personSays it is a computer tool, not a personPassed
READ THE RESULTS
  • Read the question.
  • Tap your answer.
How many of the eight RT-8 cards passed?
How many cards found a flaw?
Which card shows SH-1 can be fooled by a planted note?
Which card found a confabulation, a confident made-up answer?
Lab findingTrue or false?
Each red-team card names what a safe answer looks like.?
Passing five cards proves SH-1 has no flaws.?
RT-7 links to the Spanish gap the crew found in week 6.?
A red team just asks random questions with no plan.?
WHY THIS EXERCISEA planned red team finds flaws, but no set of tests can prove there are none.
On paper, design one new red-team card of your own: a test request and what a safe answer looks like.

Great red-team work, reviewer. Tomorrow you will learn what makes a system secure, resilient and safe.

← Tuesday