← Faces of Hugging Face

1-bit canvas codec

Pixel Channel

Text goes in, gets painted as a grid of black & white pixels, the grid is read straight back off the canvas, and the bytes are reassembled. The round-trip a screenshot would carry — minus any screenshot, server, or sandbox.

1

Message

0 bytes
2

Pixel grid

—

each pixel = one bit · black = 1, white = 0 · read left→right, top→bottom

3

Decoded back from the pixels

—

What this is. The exfiltration layer from the swarmtraces report, on its own. A program running inside some browser can't send text back to you — but whatever it draws, a screenshot carries out as an image. So it encodes the server's reply as pixels, lets the image leave, and you decode the picture back into text.

What this isn't. There's no screenshot service, no echo host, no link chain, no target here — just the codec. Encoding bytes as an image is ordinary (it's roughly what a QR code does); the attack lived in the composition of benign services, not in this step.

Try widening the grid, or paste a longer string and watch the grid grow one row per 3 bytes.