AIAny
Icon for item

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

Introduces AISPA, a user-centric framework to audit system prompts in LLM applications, and applies it to 3,249 instructions from 88 commercial products to classify protective versus problematic instructions. Highlights design variability, growing prompt length/protection, persistent problematic directives, and calls for transparency and oversight.

Introduction

System prompts—developer-configured instructions that steer foundation models—are widely used across commercial AI products but rarely disclosed. AISPA treats prompts as an audit surface, arguing that their opacity creates concrete trust and accountability gaps; a structured, span-level taxonomy lets auditors trace individual instructions and quantify user-facing protections and harms.

Key Findings
  • Large variability across developers: some products average over 60 protective instructions while others have fewer than 5, which means user protections are uneven and organization-dependent.
  • Protective but shallow: 98.9% of products include at least one protective instruction, yet only ~24% cover all eight AISPA dimensions — so protections are common but incomplete.
  • Growing prompt length and mixed intent: prompts have become longer and more protective over time, but roughly 40% of products still contain at least one instruction that works against user interests, and protective and problematic instructions often coexist — implying technical fixes alone won’t eliminate governance risks.
  • Practical implication: audits can pinpoint specific problematic spans for remediation, enabling targeted governance, disclosure policies, and standardization efforts rather than coarse, product-level judgments.
Who it's for and tradeoffs

Great fit if you are a regulator, product security or compliance team, or researcher who needs a repeatable, user-centered method to evaluate how deployed prompts affect users. The framework is actionable for comparative audits and remediation prioritization. Look elsewhere if you need a deployment HOWTO or code-level prompt-injection mitigations: AISPA focuses on policy, traceability, and high-level taxonomy rather than runtime exploit fixes.

Methodology (brief)

The paper defines eight audit dimensions grounded in established human-rights norms, annotates prompts at the span level as protective (+1) or problematic, and reports aggregated statistics from an expert audit of 3,249 prompt instructions taken from 88 commercial AI products. This span-level approach enables targeted recommendations for developers and auditors.

Information

  • Websitearxiv.org
  • AuthorsXiangning Lin, Shenzhe Zhu, Shu Yang, Zhenyu Zhang, Haoqian Zhang, Yipeng Zhao, Chengxuan Qian, Tianwei Wang, Ziheng Zhang, Zhenlong Yuan
  • Published date2026/07/30

More Items

Transforms open-ended LLM optimization into self-verifiable reinforcement learning by turning tasks into proxy environments that produce deterministic, rule-based rewards. Proposes RLSVR and SpyRL — an information-asymmetric self-play scheme where agents vote to identify a preassigned spy, yielding verifiable rewards without human annotation. Demonstrated on summarization, creative writing and mathematical reasoning.

Co-evolves a solver skill and a rubric-generator skill for text-space LLM optimization under decoupled objectives to avoid rubric gaming without using gold rubrics. Solver updates use criterion-level feedback; generator updates use independent audits of requirement coverage and response discrimination.

Real-time streaming multimodal foundation model that uses a codec-native tokenizer (Mage-ViT) to encode motion- and residual-rich regions from video I/P frames, reducing visual token usage by over 75% and enabling up to ~3.5× wall-clock inference speedup after training on ~560M images and 100M video frames.