Mozilla: ChatGPT Can Be Manipulated Using Hex Code
LLMs tend to miss the forest for the trees, understanding specific instructions but not their broader context. Bad actors can take advantage of this myopia to get them to do malicious things, with a new prompt-injection technique.