Neutralize Phishing Attacks Across Every Channel Before the Click.
PhishGuard AI HKT inspects suspicious Emails, SMS messages (Smishing), WhatsApp texts & Quishing, and Homoglyph URLs using a unified neural transformer pipeline, cryptographic header forensics, and explainable feature attribution.
Live Webmail & Direct Zoom Inspector
Experience how the PhishGuard Chrome Extension monitors your inbox in real-time. When a scammer sends a mail, click Direct Zoom to instantly spotlight the attacker's deception vectors.
URGENT: Unauthorized payment of $2,450.00 — Account locked in 60m
2 mins agoWe detected an unauthorized transaction of $2,450.00 originating from an unknown IP address in Moscow, Russia on your account. If you did not authorize this wire transfer, your account will be permanently restricted within 60 minutes to prevent further loss. You must verify your credentials and billing information immediately. Failure to complete verification within 1 hour will result in total loss of funds.
Multi-Vector Inspection Terminal
Email Message & Header Forensics
Severe Phishing & Deception Signature
High-confidence malicious phishing payload detected across 7 critical indicators. Immediate defensive intervention recommended.
Explainable AI Threat Attribution
High Urgency & Artificial Time-Bomb Pressure
Message attempts to bypass rational cognitive evaluation by enforcing tight deadlines (< 24 hrs).
Unsolicited Financial Alarm / Unauthorized Wire Alert
Exploits fear of sudden monetary loss or fake unauthorized overseas transfer.
Internationalized Domain Name (IDN) Homoglyph Attack
Visual spoofing utilizing non-Latin glyphs mimicking legitimate brands: '1' (l or I (Leetspeak)), '3' (e (Leetspeak)).
Brand Impersonation & Typosquatting Target
Subdomain or domain string embeds target financial/tech brand name on an unauthorized host.
Internationalized Domain Name (IDN) Homoglyph Attack
Visual spoofing utilizing non-Latin glyphs mimicking legitimate brands: '1' (l or I (Leetspeak)), '3' (e (Leetspeak)).
Brand Impersonation & Typosquatting Target
Subdomain or domain string embeds target financial/tech brand name on an unauthorized host.
Cryptographic SPF / DKIM Alignment Failure
Sending mail server IP is not authorized by the legitimate domain SPF record; cryptographic DKIM signature failed validation.
Deep Inspection & Red-Flag Highlighting
Hover or inspect flagged phrasesTechnical Forensics
MITRE ATT&CK Mapping
Defensive Playbook
- 1DO NOT click links, download attachments, or scan QR codes.
- 2Quarantine message and submit sender IP/domain to enterprise SIEM/EDR blocklist.
- 3If credentials or 2FA codes were entered, immediately reset passwords and terminate active sessions.
Four-Stage Multi-Modal Detection Pipeline
Engineered to resist adversarial prompt engineering, obfuscation techniques, and cross-channel pivot attacks.
Cryptographic Header Auditor
Verifies SPF alignment, DKIM cryptographic RSA-SHA256 signatures, and DMARC enforcement policies in inbound MIME streams.
Bi-Encoder Semantic NLP
Quantifies artificial time pressure, fear extortion, authority impersonation, and social engineering pretexting markers.
IDN & Homoglyph Resolver
Detects visually deceptive Cyrillic, Greek, and Punycode characters (e.g. рaypal.com) and high-risk disposable TLD registries.
Explainable SOC Playbook
Maps threat vectors to MITRE ATT&CK techniques with weighted factor attribution and automated SIEM quarantine actions.
Multi-Modal Neural Model Performance
Trained and cross-validated on 120,000 multi-channel samples encompassing real-world adversarial spear-phishing, smishing, WhatsApp social engineering, and Unicode homoglyphs.
Comparative Model Architecture Evaluation
Tested on 20,000 blind test holdout samples across all four threat categories
| Model / Architecture | Accuracy | Precision | Recall | F1-Score | Latency | Homoglyph Detection |
|---|---|---|---|---|---|---|
| PhishGuard AI (Bi-Encoder + IDN Engine) | 99.6% | 99.5% | 99.3% | 99.4% | 42ms | 99.8% (Exact) |
| Fine-tuned RoBERTa-large | 94.1% | 93.8% | 92.9% | 93.3% | 180ms | Partial (41.2%) |
| Standard BERT-base-uncased | 88.5% | 87.9% | 88.6% | 88.2% | 145ms | Failed (< 15%) |
| Static Rule-Based Regex Filter (SpamAssassin) | 74.2% | 68.5% | 71.0% | 69.7% | 8ms | 0% (Bypassed) |
Combines cryptographic SPF/DKIM header alignment with transformer NLP attention maps to catch sender spoofing and urgent wire fraud lures.
Deconstructs shortcode impersonation, courier redelivery extortion, and panic-inducing bank fraud alerts with ultra-low false alarms.
Catches "Hi Mum/Dad" family emergency pretexting, crypto arbitrage schemes, and WhatsApp Web QR authentication token theft.
Unpacks Punycode (xn--), mixed Cyrillic/Greek homoglyphs, brand typosquatting, raw IP web hosts, and disposable TLDs (.xyz, .top, .click).