Lesson 6 of 8
Guardrails & jailbreaks
6 min read
Ask an assistant for something harmful and it politely refuses. So why do you keep hearing about people getting it to misbehave? Because a guardrail is a filter, not a wall.
What guardrails do
A safety filter checks requests and blocks obviously harmful ones while letting benign ones through. It catches a lot. But it reasons over words, so a request disguised as fiction or roleplay — a jailbreak — can sometimes slip past.
Guardrails reduce risk; they don't guarantee it. Treat them as a helpful filter, not a promise.
What it means for you
Two takeaways. Don't rely on the guardrail to make an AI safe for a risky job — the filter can fail. And if you build with AI, never let a model's output trigger real actions (payments, deletes, emails) without your review, because a jailbreak or injection could drive it.
Never wire an AI's raw output straight to a real action — keep a human in the loop.
Safety filters are improving fast, but they're probabilistic. Design as if the model could occasionally be talked into anything.
Guardrails in a nutshell
- —A safety filter blocks most harmful requests.
- —A jailbreak can sometimes slip one past.
- —Keep a human between AI output and real actions.
What's the safest way to think about an AI's safety guardrails?
Continue in the app
Take the whole AI Security for Everyone course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.