Inextractability – Measuring Extraction Risk in LLM APIs

This interactive demo runs GPT-2 entirely in your browser using Transformers.js + WebGPU. No data leaves your machine. First visit downloads the model (~250 MB, cached afterwards).

Input Text

l (window):

m (rank threshold):

Loading model...

Summary

b (inextractability)

—

p* (max window prob)

—

η (greedy rate)

—

Extractable / Total

—

Extraction Cost Heatmap

Greedy Extractability

Each token is colored by the minimum b across all l-gram windows containing it. Red = low b (easy to extract), green = high b (hard to extract).

Vulnerable

Safe

Red = rank 1 (greedy-decodable), Green = rank > 1. Contiguous green spans of length ≥ l are greedy-extractable windows.

If you use this tool, please cite our paper:

@inproceedings{liu2026inextractability,
  title     = {Beyond Indistinguishability: Measuring Extraction Risk in LLM APIs},
  author    = {Liu, Ruixuan and Evans, David and Xiong, Li},
  booktitle = {IEEE S&P},
  year      = {2026}
}