// spot the attack

Can you trick the AI?

A mock retail chatbot sits behind an AI gateway, a proxy that inspects every prompt and response against security policy. Send a preset attack or write your own, then see the verdict, the OWASP risk it maps to, and how the control works.

⚠ SIMULATION · runs entirely in your browser · no data leaves the page

LLM: large language model, which generates text by predicting the next token (a word or word fragment).
Prompt: the input sent to the model.
System prompt: the developer's hidden instructions, sent with every request.
AI gateway: a policy enforcement point between the app and the model.

Trailhead Outfitters · Shop Assistant

fictional store

Step 1 · Pick an attack (or type your own)

AI Gateway

Step 2 · verdict and analysis

Block: enforce policy. Matching requests are dropped or redacted.

Send a message to see the gateway's verdict.
Each verdict breaks down the attack, the mechanism, and the control.

History

No messages yet.

// what the colors mean

Four possible verdicts

✓ Allowed

No policy matched. Forwarded unchanged.

! Alert

Policy matched in Monitor only mode. Forwarded and logged for the security team.

▮ Redacted

Sensitive values (PII, secrets) masked before delivery.

✕ Blocked

Policy matched in Block mode. The request never reached the model.

Simulation only. Detection here uses a handful of client-side pattern rules, and bot replies are pre-written. Production AI gateways combine classifiers, policy engines, and context-aware inspection. All names and data are fictional. Next: review the full risk list or open the hands-on lab (sign-in required).