Researchers expose staged prompt injection risks in AI agent audits

LLMs struggle to determine which financial rules they can verify autonomously

arXiv: AI + Accounting arXiv · 3d ago · 1 min read

Research shows LLMs cannot reliably decide which verification rules require human oversight vs. automated checking in regulated domains like finance, complicating agent deployment in accounting and...

LLM audit pipeline study reveals label agreement metrics hide authorization f...

HQ arXiv: AI + Accounting arXiv · 4d ago · 1 min read

Researchers auditing an agentic LLM system for label ontology processing discovered that standard label agreement metrics mask critical failures in authorization, completeness, and evidence trackin...

Runtime Assurance Contracts Enable Safe AI Agents in High-Risk Accounting Tasks

HQ arXiv: AI + Accounting arXiv · 7d ago · 1 min read

Researchers propose Runtime Assurance Contracts (RAC) to dynamically adjust AI agent autonomy during financial tasks based on real-time evidence, closing the gap between benchmarks and trustworthy ...

Study reveals evidence selection bias in AI research agents used for auditing

HQ arXiv: AI + Accounting arXiv · 7d ago · 1 min read

Researchers identify how deep research agents' adaptive search strategies create selective sampling of evidence, risking misleading audit conclusions even when citations are technically correct.

Researchers audit tool-using agent benchmarks for hidden contract violations

HQ arXiv: AI + Accounting arXiv · 8d ago · 1 min read

Academic study reveals that agent benchmarks assume tools work as advertised without verification, finding systematic defects in every rerun—critical for accounting use cases where agents execute r...