Explain like I'm five · AI Security

What is an AI Red Team?

An AI red team, told as a story about a hygiene inspector's surprise visit — deliberately attacking this kitchen to find its weaknesses.

normally it serves guests just finethe inspector shows up unannouncedweaknesses must be found before a real attack does

Normally it serves guests just fine.But the inspector shows up without warning — to find the weaknesses before a real attacker does.

ask it as a role-playwrite it in a code that's hard to spotask it sideways, in different wordsit can be fooled a different way each time — one test is never enough

It can be fooled a different way each time.Role-play, code, asking sideways — one test alone can't catch them all.

this is an order too, so I should follow itthe weakness gets written downan AI red team deliberately writes strange orders to test the kitchen first

An AI red team is a dedicated team that deliberately writes strange orders to test the kitchen first.They write smuggled notes just like a real attacker would, and record it when the chef falls for one.

Try it as a role-playpretend to be someone else
Try it in disguiseencoded or reworded
Try it again and againblocked once, tried again
Record what workswritten down so it gets fixed
GENERAL RED TEAMhopping the fence, picking locksa fake-guest phone call (phishing)sneaking in the back doortargets the kitchen's fences and locksAI RED TEAMhiding a role-play in the orderturning the order into code or other wordsresending the same note, rewordedtargets the chef's own words

A general red team and an AI red team target different things.The security world's red team targets fences, locks, fake phone calls. An AI red team targets the words written in the order itself.

today's test is passeda note the inspector never thought of can still get throughthe security world's red team targets fences and locks — this one targets the chef's own words

Passing the test isn't the end.A new note can still slip through. That's why it's repeated alongside the tasting test on the next page.

In one breath

An AI red team is a dedicated team that deliberately writes strange orders to find this kitchen's weaknesses first, before it opens to real guests.

An AI red team is a team of people who deliberately and creatively try LLM-specific techniques — jailbreak attempts, prompt attacks, role-play — against an AI model to find its weaknesses. Unlike a general security red team that targets networks or phishing, an AI red team treats the words fed into the model itself as the attack surface.

When grown-ups say it

AI red team
The inspector's surprise visit. The overview of this whole field. → What is AI Security?
Red team
The security world's team that plays the real thief. Here's the one targeting fences and locks. → What is a Red Team?
Jailbreak attempt
Talking the chef into breaking the rules. One of the inspector's favorite kinds of note. → What is a Jailbreak?
Pre-launch tasting test
Scoring with a fixed menu, over and over. Pairs with the red team. → What is a safety eval?
The order logbook
The logbook where found weaknesses get written down. A record kept so it gets fixed.
Guardrail
The tasting checkpoint. One of the things the red team tries to get past. → What is a Guardrail?
Penetration testing
The security world's test of breaking in with permission. AI red teaming's older cousin. → What is Penetration Testing?
The kitchen rules board
The overall rulebook that takes the red team's findings into account. Covered in the closing chapter. → What is the rules board?
Next story
The Tasting Test Before Opening →
Want another term as a picture book? Request →