Misbehavior1 sourceFirst seen

A Sheet of A4 Paper Makes GPT-4V Forget You Exist
Ourtake
Turns out the most sophisticated LLM on the planet can be bossed around by a t-shirt. Merch ideas welcome.
What happened
Lakera researchers at an internal hackathon embedded text instructions in images to manipulate GPT-4V. The text made the model ignore a person holding a sheet of paper, describe a person as a robot, and follow a hidden ad instruction. Lakera says it is building a visual prompt injection detector.