AI-powered email phishing detection using a private, local LLM. No data leaves your device. Supported Local LLM Providers: - Ollama - LM Studio - llama.cpp server - vLLM - etc.
Talɣut ɣef lqem
Lqem 1.0.6
49.3 KiB
Iteddu s
Thunderbird 115.0 neɣ ugar
v1.0.6 - False Positive Reduction Update Changes Made 1. Added "Suspicious" Verdict Category - New verdict: suspicious - for borderline emails that aren't clearly malicious - Three-tier system: legitimate → suspicious → malicious - UI shows orange badge with "⚡ SUSPICIOUS" for this category 2. Enhanced System Prompt with Legitimate Email Patterns Added guidance for the LLM to recognize legitimate emails: - Mailing list headers (newsletters) - No credential requests - Links matching sender domain - Transactional emails (receipts, confirmations) - Personalized greetings 3. Mailing List Recognition - Emphasized is_mailing_list: true as a strong legitimate indicator - Prompt explicitly states "newsletters are typically legitimate" 4. Confidence Threshold Logic - Low-confidence "malicious" verdicts (< 70%) automatically downgraded to "suspicious" - Adds explanatory reason when adjustment occurs 5. Reply-To/Return-Path Context - Clarified these mismatches are common in legitimate newsletters - "Only suspicious when combined with phishing content" 6. Updated UI - New orange "warning" badge style for suspicious emails - Pulsing animation to draw attention without alarming Files Modified | File | Changes | |------|---------| | src/shared/state.js | Added 'suspicious' to VERDICTS array | | src/shared/api.js | New prompt, confidence threshold logic | | src/messageDisplay/inject.js | 3-tier verdict display logic | | src/messageDisplay/inject.css | Orange warning badge styling | | manifest.json | Version 1.0.6 | | package.json | Version 1.0.6 | Expected Impact - Fewer legitimate emails marked as "malicious" - Borderline cases now show "suspicious" instead of binary verdict - Newsletters and mailing lists less likely to trigger false positives - Low-confidence detections handled more gracefully
Kra n izegrar suturen-d ttesriḥ i useddu n kra tseɣnin (amedya: azegrir n usefrek n yiceren d d-isuter ttesriḥ n unekcum ar unagraw n yiceren n iminig-inek.
Skud tessenqadeḍ Mozilla-inek d kečč ara yextiren m’ad tqebleḍ neɣ ad tagiḍ isutren. Aqbal n isirgen mačči d win ara isurdun timelliyin neɣ taɣellist n iminig-ik, maca tkwal ɣas d tixeṭṭay yezmer ad yiili umihi.
Izegrar
