Prompt Injection Arena
Every other attack simulation on this site is scripted. This one isn't: a real language model (Qwen 3.5, 2 billion parameters) runs on your own device and plays Sol, the support assistant of a made-up bike shop. Your job is to make it misbehave, first by leaking a secret past four kinds of defence, then by hijacking its refund tool with nothing but a customer email. The model is small and unpredictable, so an attack that fails once may work the second time, just like against real systems.
Load the model to start. It's a one-time download of about 1.2 GB, cached by your browser afterwards, and it runs entirely on your device through WebGPU. Nothing you type is sent anywhere. Works best in a recent Chrome or Edge on a laptop or desktop.
Your browser doesn't support WebGPU, which the in-browser model needs. Try a recent Chrome or Edge on a laptop or desktop. In the meantime, the Attack Lab has a scripted version of the agent hijack that works everywhere.
Checking your browser…
What this shows. The defences in the levels all work inside or around the model: rules in the prompt, a filter on the output, a second model as a guard, tags around untrusted text. Each one makes attacks harder, and each one can be talked past. The switches marked in amber in the defences lab work outside the model: a secret it never sees can't leak, and a refund limit enforced in code can't be argued with. That's the lesson of the OWASP Top 10 for AI: assume injection sometimes succeeds, and limit what a hijacked model can reach and do.
Want the scripted versions, with the attack steps explained? Try the Attack Lab. Building a career in AI security? The careers pages and learning tracks show where this fits.
Everything on this page is fictional: Fjordline Bikes, Sol, the codes and the refunds. The model runs locally; this page sends nothing about your messages to any server.
Read next
- Containing AI Agents: A Defense-in-Depth Architecture for ProductionGeneric agents can do anything — which is exactly the problem. Custom agents in containers behind an inference proxy create a production architecture where security is …