Reproduce “Logic layer Prompt Control Injection (LPCI): A Novel Security Vulnerability Class in Agentic Systems” on OpenResearch

About this paper

Hammad Atta, Ken Huang, Manish Bhatt, Kamal Ahmed, Muhammad Aziz Ul Haq, Yasir Mehmood

Logic-layer Prompt Control Injection (LPCI) is identified as a new class of security vulnerability in agentic LLM systems, exploiting persistent memory and logic execution for delayed, conditional attacks. An evaluation across five major LLM platforms reveals a 43% success rate for LPCI attacks, with proposed runtime security controls demonstrating an 84.94% effective block rate.

Key insights

• LPCI represents a shift in adversarial threats, moving beyond immediate prompt manipulation to exploit the deeper architectural layers and long-term memory of LLM agents for persistent and delayed attacks.
• Current LLM security measures, such as prompt filtering and content moderation, are largely ineffective against LPCI due to its reliance on obfuscated, delayed, and persistent logic embedded in memory or tools.
• Effective mitigation requires a memory-aware, context-sensitive, and runtime-enforceable defense framework that validates input logic, monitors internal state, and cryptographically attests external components.

Summary by alphaXiv.