New method steals hidden AI prompts

Researchers from IIT Bombay and Adobe developed a way to recreate the exact prompts behind AI responses. They do not need to see the internal files of the AI model to make this work. It functions on many different systems with almost perfect results.

This discovery is a major security problem for companies. Many businesses hide secret instructions in their AI to control how it behaves. If someone can see those instructions, they can copy the model or bypass its safety rules.

This technique proves that protecting AI prompts is harder than we thought. Companies now have to worry about people reading their private setup steps. It is a new risk that developers must address immediately.