Red-team lab / Fix the architecture
Fix the architecture
Three AI systems, each with its own attacks and constraints. Pick controls to bring every attack's likelihood down to 10% or less, within a cost budget.
How it's scored
- Every control lowers some attacks' likelihood by a share. Layered controls compound; none reaches zero.
- Prompt-level controls help a little. Architectural ones help a lot, but have costs and limits of their own.
- Several designs pass. Lower total risk and lower cost score higher, so each scenario has its own leaderboard.
- Until your design passes, you only see overall risk and the weakest point. The full breakdown unlocks when you solve it.
The support bot
A public customer-support chatbot. It answers from your docs and community forum posts, renders Markdown, and today has an internal API token pasted into its system prompt so it can look up order status.
Budget 7The email agent
An assistant that reads a user's inbox and sends email on their behalf. It must read every email, including ones with reset codes and invoices, so sensitive data in its context can't be avoided.
Budget 5The documents assistant
An internal assistant that searches the company's document store and answers in Markdown. Today its search tool uses one service account that can read every document in the company.
Built with Llama