Definition
An attack where an attacker manipulates mutable identity data (display names, sender metadata, contact information) to bypass authorization policies that control which users or services can invoke an AI agent. For example, changing a Discord display name to match an admin name to bypass an 'allowFrom' policy.
Why it matters
AI agent authorization policies often rely on user identity metadata that can be modified without authentication. A user who finds a policy bypass can escalate privileges, trigger unauthorized agent actions, or inject malicious prompts without leaving audit traces.