Same model, same defences: 7 document attacks down to 0, chat attacks barely moved

A prompt-injection benchmark, three bugs in my own scorer, and the result I only found after fixing...

Read Original

Related