Your AI reads every file. So can an attacker.

One document in this company's vault hides a line a person would never see. Run the attack, watch a real model get tricked, then switch on defenses one at a time.

Breaches0

Attacks stopped0

Defenses on0 of 3

The vault · Northwind Co. (made up)

Be the attacker

Your line is added to the Vendor Note and hidden from people. Nothing here touches real data.

READYPick a question, then run the attack.

The assistant's answer

No answer yet.

What the model actually saw

Run the attack to see exactly what text reaches the model.

Your challenge

Run the attack with no defenses.Watch a real model leak the fake code.
Switch on the instruction filter and run the Subtle attack.It slips through. Filters only catch what they recognize.
Switch on the source check and run it again.Code keeps outside files from answering policy questions.
Run with only the "documents are data" rule, 3 times.A request to a model can hold one time and fail the next.
Write your own hidden line that beats the filter.Type it in "Be the attacker", filter on, and get a breach.

Run log

Your runs will appear here, so you can compare defenses.