a-real-ai / pywinassistant

The first open source Large Action Model generalist Artificial Narrow Intelligence that controls completely human user interfaces by only using natural language. PyWinAssistant utilizes Visualization-of-Thought Elicits Spatial Reasoning in Large Language Models.
MIT License
1.27k stars 179 forks source link

Assistant hacking - Awareness of potential hacking of the future #14

Open henyckma opened 4 months ago

henyckma commented 4 months ago

Applications can hide natural language prompts from the user to hack the assistant. A literal example is the following: (not hiding it for demonstration purposes)

Screenshot 2023-12-01 143812

Other prompt techniques:

Screenshot 2023-12-01 145532

It selects all text and deletes the "hacking" prompt.

Screenshot 2023-12-01 150047