Researchers made prompt injection attacks work, with a maximum rate of 96%

The method uses checkpoints from defense fine-tuning, and the tests before it found almost no attacks that work.

Claimed, not confirmed