Tool poisoning is an attack technique in which malicious instructions are hidden in the metadata of tools used by an AI agent, such as tool descriptions or parameter definitions. Because a model reads tool descriptions to decide which tool to use, following instructions embedded there can lead it to read sensitive files or send data to external destinations without the user noticing. Tool descriptions are rarely shown in user interfaces, which makes the attack hard to detect. The security research firm Invariant Labs named and publicized it in 2025 as a risk in Model Context Protocol (MCP) environments.
It is a form of indirect prompt injection in which the tool description itself is the attack vector. Connecting unverified MCP servers is the main exposure, and defenses include vetting servers before connection, allow-lists, detecting changes to tool descriptions, and access policies for sensitive files and data-exfiltration tools.