Data Leakage Prevention in Agentic Applications via Preemptive Hardening

Abstract

Agentic systems integrate LLM driven planning with interfaces to external tools, making data leakage and tool misuse feasible via instruction/data boundary failures and prompt injection attacks. Enforcing required controls consistently is particularly challenging in workflows spanning many codebases and heterogeneous agents. To address this challenge in multi agentic systems, we present a pre-deployment pipeline for scanning, hardening, and validation of agentic applications. The pipeline analyz

Categories

Framework mappings

Suggested from the entry's categories.

@misc{shukla2026data,
  title = {{Data Leakage Prevention in Agentic Applications via Preemptive Hardening}},
  author = {Akansha Shukla and Emily Bellov and Parth Atulbhai Gandhi and Yuval Elovici and Asaf Shabtai},
  year = {2026},
  month = jul,
  eprint = {2607.18847},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.18847}
}