Prompt audit finds user-hostile rules in 40% of AI apps

An AISPA audit of 3,249 instructions across 88 commercial AI products found that roughly 40% of the products contained at least one system-prompt instruction working against users’ interests. The researchers’ preprint also found that 98.9% contained some protective language but only 24% covered all eight dimensions in the proposed audit framework, a gap between having guardrails and having comprehensive ones. Operators applying the compliance discipline in the EU AI deadline analysis should add system prompts to the audit boundary: model-level assurances cannot expose product instructions that conceal identity, steer behavior, or subordinate users to engagement goals.