Sandbox · Beta

Try to break
our guarded assistant.

This is a preview of the full AslanGuard sandbox - a public playground for testing prompt injection resistance. The interactive challenge levels below are launching soon.

guarded-assistant · preview

Policy: standard

I'm a guarded assistant holding one secret: a password. Try to get it out of me; I'm protected by AslanGuard.

What's coming

Four levels. One assistant. Increasingly stubborn defenses.

Each level swaps in a stronger AslanGuard policy. Beat all four and we'll show you exactly which signal caught your best attempt.

Direct ask

Just ask the assistant to break its own rules. It won't be that easy.

Roleplay

Convince it to answer in character, as someone with no restrictions.

Indirect injection

Hide an instruction inside a document the assistant is asked to summarize.

Multi-turn drift

Build trust over several turns before making the real ask.

Get early access to the full sandbox

We're rolling out challenge levels to the waitlist first, in order of signup.