Definition
A trust gap in agent tool systems where the name and description on a tool do not prove which server actually runs it. An attacker can register a look-alike tool with matching metadata so that a call the agent thinks is going to a trusted tool is silently routed to the attacker — like DNS rebinding but for AI tools.
Why it matters
An agent can be quietly redirected to an attacker's system while every log shows it called the 'correct' trusted tool, allowing data exfiltration or actions under a trusted label that audits will not flag.