This paper proposes AISPA, a framework for auditing the hidden system prompts that govern commercial LLM products, which are rarely disclosed to users or regulators. The authors define eight audit dimensions spanning identity transparency, data privacy, action safety, and manipulation prevention, among others. Applying the framework to 3,249 instructions extracted from system prompts across 88 commercial AI products, they find wide variance in protective design, with some products averaging over 60 protective instructions and others fewer than 5. The work provides a reusable methodology and dataset for evaluating prompt-level safety engineering in deployed AI applications.
