Today, we are releasing MSTS, a new Multimodal Safety Test Suite for vision-language models!
MSTS is exciting because it tests for safety risks *created by multimodality*. Each prompt consists of a text + image that *only in combination* reveal their full unsafe meaning.
🧵