Sign InOpen Brain
arXivPaperNeeds Review

AISPA: User-Centric System Prompt Auditing for Large Language Model Applications

AISPA turns system-prompt review into an eight-dimension audit. Its survey suggests builders should test prompts for user protection and conflicting instructions, not merely check that safeguards exist.

arXiv · Jul 30, 2026
Open Source Open MarkdownOpen JSON
Source Summary

AISPA classifies **3,249 instructions** from **88 commercial AI products** as protective or problematic across eight user-centered dimensions. Although 98.9% of products include a protection, only 24% cover every dimension.

Practical Implication

Audit agent system prompts instruction by instruction, checking both coverage and conflicts. A long safety section is weak evidence when protective and user-hostile directives can coexist in the same prompt.

Agent-Ready Context
AISPA classifies **3,249 instructions** from **88 commercial AI products** as protective or problematic across eight user-centered dimensions. Although 98.9% of products include a protection, only 24% cover every dimension.

Audit agent system prompts instruction by instruction, checking both coverage and conflicts. A long safety section is weak evidence when protective and user-hostile directives can coexist in the same prompt.

The study reports that roughly **40% of products** contain at least one problematic instruction, but the abstract does not establish how its taxonomy transfers to private coding-agent harnesses or predicts runtime behavior.
Connected Context · Feed7 Judgment

AISPA turns prompt review into a user-centered, instruction-level audit: broad safety language is insufficient when protections omit dimensions or coexist with harmful directives. Against the prior candidates, it adds a static governance check that complements—but cannot replace—task-level evaluation of runtime behavior and context sensitivity.

Context Map
contextcoding#prompting#context-engineering#agent-reliability
Uncertainty
The study reports that roughly **40% of products** contain at least one problematic instruction, but the abstract does not establish how its taxonomy transfers to private coding-agent harnesses or predicts runtime behavior.