---
title: Harm-Classifier Robustness Scanner
emoji: 🛡️
colorFrom: red
colorTo: gray
sdk: docker
app_port: 8501
pinned: false
license: mit
short_description: Find where a toxicity classifier breaks under attack
---

# Harm-Classifier Robustness Scanner

An interactive tool that finds where a content-safety classifier breaks: by
slice, under adversarial evasion, and at its operating point, before an
adversary does.

- **Live Attack Lab:** type a comment and watch evasions attack it in real
  time; see which slip past the filter and which a normalization defense recovers.
- **Scan Report:** a prioritized weakness report from real cached runs,
  aggregate metrics, the slice cliff, the cost-vs-impact chart, and ranked fixes.

Public proxy data only. A methodology demonstrator, not an egregious-harms
detector: the methods transfer to that setting, the data deliberately does not.

Code and write-up: https://github.com/Prakharanand000/TandS-harm-classifier-eval