What happened
On Oct 8 AegisAI (SF startup from ex-Google reCAPTCHA/Safe Browsing leads) announced Proteus, a purpose-trained foundational model for email security that runs a per-message structured investigation — sender infrastructure/domain history, following embedded links, and reasoning about the intended action — instead of signature/rule matching. AegisAI reports >99% detection on 3,020 human-reviewed phishing attacks versus 81%/80% for unnamed closed/open general-purpose models.
Why it matters
A from-scratch, task-specific security model (not an LLM wrapper) targeting the weakest spot of legacy SEGs — novel, AI-crafted phishing and BEC with no historical signature. Founders' abuse-detection background (reCAPTCHA, Safe Browsing, Web Risk) lends credibility, though benchmark is self-published with no FP rate or model spec yet.
Applicability
Email-security buyers and MSSPs evaluating AI-native alternatives to legacy SEGs should monitor Proteus for independent/FP-rate validation before adoption; start with internal evaluation of novel-phishing and BEC detection.