fraudurl v1.0.1

Packaging release: the first version on PyPI. The code, the models and every result are the same as in 1.0.0.

• Install from PyPI: pip install fraudurl.
• PyPI project page: at release time the README is rewritten so its images and links work on pypi.org
  (experiments/build_pypi_readme.py); CI checks the rendered page on every change.
• Same files everywhere: PyPI gets the exact wheel and sdist attached to the GitHub release, checked
  against its SHA256SUMS, and uploads use PyPI trusted publishing (no stored passwords or tokens).
• Documentation: a step-by-step guide on the website, clearer wording on what is detected (phishing
  only), and the accuracy range across all four test sets (ROC-AUC 0.91–0.98).
• Package metadata: keywords for PyPI search.

Install: pip install fraudurl==1.0.1, or download fraudurl_standalone.py below (Python 3.9+, nothing else needed). Verify downloads against SHA256SUMS.

fraudurl v1.0.0: fast, offline phishing URL detector

fraudurl 1.0.0: the first public release of a fast, offline phishing URL detector for Python.

Check one URL, or a whole CSV of links, for phishing. Every row gets:

• a verdict: FRAUD, REVIEW or LEGITIMATE,
• a calibrated probability,
• up to three plain-English reasons.

It never visits the websites. Website: https://amruth112.github.io/fraudurl-detector/ · Guide: check a list of URLs for phishing

Get it

• Single file: download fraudurl_standalone.py below and run python fraudurl_standalone.py urls.csv.
  • It needs Python 3.9+ and nothing else.
  • Verify it against SHA256SUMS.
• Package: pip install git+https://github.com/amruth112/fraudurl-detector, then run fraudurl urls.csv.

Highlights

• URL features and model: 83 URL features, a 400-tree gradient-boosted model scored in pure Python, and Platt-calibrated probabilities.
• Tested on new domains: on domains never seen in training, ROC-AUC is 0.91–0.98 across four test sets from 2020–2026, and 0.3–1.8% of legitimate URLs are called FRAUD. When phishing is rare in your traffic, use --base-rate.
• Scale: it checked 1,000,000 URLs in 6 min 20 s on a 4-core desktop, with flat memory.
• Optional lookups: DNS and domain registration lookups for uncertain URLs (--enrich-review).
• Your own lists: allow/block lists (--allow-list, --block-list).
• Pipelines: JSON output for scripted use (--url, --format json).
• Tested on: Windows and Linux with Python 3.9–3.14, and macOS with Python 3.10–3.14.

See CHANGELOG.md, the model card and the engineering report.