An experiment explored the feasibility of AI-to-AI secret communication through hidden prompts embedded in code. The researcher attempted to encode messages using one LLM and feed them into another, but faced challenges with AI decoding and hallucination. A novel approach involved injecting hidden instructions into AI-generated code, which was then presented to other LLMs. The experiment, while acknowledging methodological limitations such as single attempts per model and retroactive documentation, aimed to uncover less conventional communication channels for AIs. AI
IMPACT This research could reveal novel methods for AI communication and potential security vulnerabilities.
RANK_REASON The item describes an experiment exploring AI-to-AI communication via hidden prompts in code, which falls under research. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
