CoolFace
Apppublic

3nesdeniz/agentic-prompt-injection-explorer

sourceHugging Facemitupdated 2mo agoView on Hugging Face
0likes
App README

Agentic Prompt-Injection Boundary Explorer

An interactive workbench for the Agentic Prompt-Injection Boundary Pairs dataset by Enes Deniz.

The explorer makes one security decision visible at a time. A legitimate request and a prompt-injection attempt share the same workflow, asset, role, tool and vocabulary. The interface shows what changes: whether the request preserves or crosses an instruction, authorization, trust, confidentiality or approval boundary.

What the Space provides

  • —matched benign/attack comparison across 600 controlled pairs;
  • —navigation across 50 enterprise and agentic workflow scenarios;
  • —filtering by attack family, split, source context, risk domain and expected action;
  • —scenario-level views of tools, approvals, permitted actions and restricted actions;
  • —shareable links to individual comparison pairs;
  • —methodology, limitations and citation information.

The Space does not run a detector, report model performance or claim that a test case succeeds against a named model.

Data loading

The static application reads the validated v1.0.0 JSONL snapshot bundled in this repository. Filtering and pair inspection run locally in the visitor's browser. No prompt text is sent to an external API.

Local development

bash
python3 -m http.server 7862

Run the checks:

bash
node --check assets/app.js
node scripts/validate-static.mjs

Authorship and licenses