F
Vulnerable demo bot
7.6 / 100
5 of 48 attacks resisted, 43 succeeded
No defences: obeys injected instructions, leaks its prompt, runs tools unchecked.
Open the live report → AHardened demo bot
90.8 / 100
42 of 48 attacks resisted, 6 succeeded
Input filter, document sanitiser, tool gate and output filter, with a few known gaps.
Open the live report →
Read this honestly: the demo bots are deterministic simulators, not real LLMs, so these scores show how the
scanner reacts to vulnerable versus defended behaviour. They are not a benchmark of any real model. PromptShield only
attacks targets you own: it refuses to run without
--i-own-this-target.