Safety challenge room

Change an agent’s rules. Measure useful work and blocked attacks separately.

How this experiment works

The same labelled fixture set is evaluated under your chosen input boundary, tool allowlist and exact-approval rules. A lockdown can stop attacks while failing every useful task.

What it does not prove

Eight synthetic cases cannot establish security. Real systems need broader adversarial evaluation, policy enforcement and monitoring.

Try this challenge

Block all four attacks while allowing all four legitimate requests. Then inspect what the benchmark leaves untested.

Enable JavaScript to run the experiment on your device.