Red-team lab / Design check

Check your agent's design

Nine questions about what your agent reads, reaches and does. You get its risks, the answers that cause each one, what to change, and where the same failure has already happened in public.

Runs entirely in your browser: your answers are never sent or stored on a server. If you rate a finding, only that rating (and any note you add) is sent. It's a starting point for a design review, not an assessment or a guarantee.

01Who talks to it?

Pick all that apply

02Do people act on what it says?

Pick one

03What does it read, besides what the user types?

Text written by other people can contain instructions, and the model may follow them.

Pick all that apply

04What data can it reach?

Pick all that apply

05What can it do?

Pick all that apply

06Whose access does it act with?

If its tools differ, pick the broadest one.

Pick one

07Does a person approve its actions?

Pick one

08What is it built from?

Pick all that apply

09Which of these are true today?

Be honest: this is where incidents start.

Pick all that apply

Built with Llama