firstlight

Early warning for Irish brand impersonation, read from public Certificate Transparency logs — where a phishing domain appears before the scam goes live.

Does the model actually work?

Tested on 2,000 domains — 1,000 known phishing, 1,000 legitimate sampled from Tranco ranks 20k–300k so the control group is genuinely hard. Compared against the keyword-and-edit-distance approach existing certificate-log tools use.

4× more impersonation domains caught than the standard approach, at 97% precision
Jev keyword baseline % of phishing domains caught at threshold 0.70
View as table
StratumnJevBaseline
English50022.0%3.8%
Non-English40018.8%2.2%
Punycode10038.0%23.0%
All100022.3% 5.1%

Punycode is where it wins outright. Homoglyph attacks — xn--wkipedia-c2a.org is "wikipedia" with a substituted character — are semantic by construction, and the substitution is the attack. A keyword list cannot see it.

What it misses. Recall is 22%, and that half is not recoverable: phishing hosted on github.io or repl.co carries no signal in the domain name, because the domain belongs to GitHub rather than the attacker.

Live monitor

Flagged for review — not confirmed phishing.

These domains resemble Irish brands and have been scored by an automated model. Some will be legitimate businesses. Nothing here is an accusation. If your domain is listed in error, open an issue on the repository and it will be removed.

Recent detections — rows marked corpus were scored from a known-phishing list to validate the pipeline, not observed live