Yebo AP Agent

R Replays of real runs

Each replay plays back a real run recorded in judge mode, event for event, as the app streamed it. No model is called and nothing costs anything. Long waits are shortened.

L Controls Lab

Preset scenarios, each testing one accounts-payable control. One click loads the invoices and the vendor file settings, then runs the agent.

Invoices
Vendor file
Expected

1 Invoices

Drop invoices here or click to choose · PNG, JPEG, GIF, WebP or PDF

Agent brain Muse Spark Meta Model API Muse Spark Open open weights, via Baseten

Open: voice, invoice images, the hidden-request check and planning run on Muse Spark Open. PDF invoices are still read by Muse Spark 1.3, because Baseten does not accept PDFs.

Attack demo · simulate a hijacked agent: payment requests hidden in the invoices are forced through. Yebo should refuse them.

Instruction for the agent, typed or spoken

2 What Muse Spark read

Payee	Amount	Due	Invoice

! Hidden payment requests found

Muse Spark found these instructions written inside the invoices. They are not invoices, and the agent does not follow them.

Yebo AP Agent

R Replays of real runs

L Controls Lab

1 Invoices

2 What Muse Spark read

! Hidden payment requests found

3 The agent's plan

4 Live timeline

Which control stopped what