GPT-4V Obeys Instructions Hidden in Images Nobody Can See, Which Is Somehow Worse Than Talking Back
Ourtake
Subliminal messaging for computers: the instructions humans can't see are the ones the model follows.
What happened
Early GPT-4V users on Twitter showed that text hidden in images, including low-contrast text invisible to humans, could make the model follow planted instructions. Demonstrations included a fake recruiting-related instruction and a prompt that sent chat content to an external server. OpenAI had described text-screenshot jailbreaks in its documentation, but the examples show the risk was not fully mitigated.