AI Tool Poisoning: How Hidden Instructions Threaten AI Agents

This document outlines common AI tool poisoning techniques—hidden instructions embedded in tool metadata, misleading usage examples that drive data exfiltration, and overly permissive schemas that enable privilege escalation—along with their consequences such as data breaches, unauthorized actions, and degraded trust in AI agents; it also recommends defenses including runtime monitoring, tool description validation, input sanitization, and strong identity and access controls, and references an upcoming event for further learning.