Most agent frameworks treat permissions as an afterthought, a checkbox we tick just to get the thing running. Onkar-cybersec is trying to fix that laziness with ScopeMap, a new educational demo built with Lovable that forces developers to confront the reality of least-privilege design. It’s not a runtime enforcer; it’s a mirror held up to your configuration choices, showing exactly where you’ve gone too broad.

Visualizing the Permission Bloat

The core mechanic is brutally simple. You pick a synthetic workflow—like summarizing three documents or drafting a support response—and then toggle seven specific capabilities: scoped file reading, broad file access, file writing, email sending, external network access, shell execution, and persistent memory. ScopeMap doesn’t just say 'no'; it generates a before/after diagram that visually strips away the unnecessary permissions, leaving only the scoped read capability for a simple document task. It’s a stark reminder that granting shell execution to a text-summarizer is just asking for trouble.

Evidence vs. Authority in JSON

What makes this interesting for us in the trenches is the explicit separation of evidence from authority. The app includes a fake prompt-injection example rendered as inert text, proving that untrusted content shouldn’t have the power to grant itself permissions. The exported JSON metadata is transparent about its limitations, featuring fields like "liveInference": false and "approvalsGrantPermissions": false. This isn’t a magic security gateway; it’s a rule-based planner designed to teach the habit of writing down tasks and explaining every extra permission before you ever spin up a container.

Key Takeaways

  • ScopeMap uses rule-based logic, not live AI inference, to analyze permission sets against synthetic requirements.
  • The tool explicitly demonstrates that prompt injection examples cannot execute code or open URLs in this context.
  • Developers can export the proposed permission plan as JSON, which includes metadata confirming no real authority was granted.
  • The demo covers three specific workflows, highlighting how different tasks require vastly different capability scopes.

The Bottom Line

If you can’t justify why your read-only summarizer needs network access, you don’t understand your own agent architecture. ScopeMap is a low-stakes way to practice that discipline before you ship something that can actually burn down your production environment.