JevMade hello@JevMade.com
← Back to experiments

Benchmarks & research

jev-phishing-bench

Source screenshot of jev-phishing-bench
SOURCE SCREENSHOTFull screenshot ↗

What it does

A phishing test that pits Jev against Claude Haiku 4.5 on 2,000 emails.

How you can use it

A developer can adapt this script to screen incoming messages. The app sends a message to Jev, an outside AI service. Instead of asking for a single verdict, it asks several specific questions together. It checks if a link uses free hosting or if the sender demands urgent action. The setup requires an access key that connects the app to Jev.

This original test uses computer-generated emails. The malicious clues are mostly in the web links and sender addresses. This does not prove the service can catch subtle tricks in mail written by a real person.

Maker-reported (not independently measured by JevMade): gives 89.5% accuracy with no fitting, and a cross-validated logistic regression on the five signals reaches 95.1%, · link, plus an eTLD+1 comparison between sender and link. The list rule alone: 91.6% accuracy,

Primitives
choice, noul
Platform
Python
Added
Project created
GitHub stars
0 (snapshot, not live)