Comprehensive Survey Maps Agentic AI Attack Surface: Tool-Call Hijacking, Memory Poisoning, Cross-Agent Privilege Escalation as Distinct Threat Classes

arXiv·medium signal

A new arXiv survey (2603.11088) formalizes the expanded attack surface introduced by agentic AI systems, cataloging threat categories that model-level security literature has systematically missed: prompt injection chains across agent boundaries, memory poisoning in persistent stores, tool-call hijacking, and cross-agent privilege escalation. The paper surveys defense mechanisms across the full agentic stack and is the first comprehensive threat taxonomy purpose-built for multi-agent architectures.