Developer Discovers AI Agent URL Permissions Work Inversely to Standard Access Models

A developer running a scheduled multi-agent system noticed inconsistent URL-fetching behavior across sessions, where some requests succeeded and others silently timed out. Investigation revealed that the underlying permission model grants access upward in a URL hierarchy — fetching a parent or root URL — rather than downward to child paths. This is the opposite of conventional permission systems like filesystem grants, OAuth scopes, and CORS paths, where access to a directory implies access to its contents. The developer confirmed the finding by deliberately testing URLs designed to falsify the initial hypothesis, a method they credit to a past costly debugging mistake. The key practical distinction is that a fetch failure due to an allowlist block and one due to a network outage look identical in logs, requiring careful experimental design to tell apart.