On October 8, 2026, AegisAI announced Proteus, a foundational AI model built specifically to investigate every email before it reaches an inbox. The launch, detailed in a company release distributed via PR Newswire, claims Proteus detected more than 99% of 3,020 human-reviewed phishing attacks in the company's internal evaluation — and positions purpose-trained AI models as the next battleground in email security.
What AegisAI announced
Proteus does not simply classify a message as spam or not-spam. Instead, it runs a structured investigation on each email, mirroring what a security operations center (SOC) analyst does by hand. The process breaks into three steps: examining sender identity (looking past the display name to the underlying infrastructure, domain history, and authentication signals), following embedded links to understand where a URL actually redirects before anyone clicks, and reasoning about intent — what action the email is trying to elicit, and whether that action fits the purported sender's relationship with the recipient.
That intent-reasoning step is the differentiator. Traditional secure email gateways rely heavily on reputation lists, malware signatures, and static rule sets. Those approaches struggle precisely where modern phishing succeeds: AI-generated lures with clean grammar, novel domains with no historical fingerprint, and first-contact business email compromise attempts that contain no known-bad signatures.
The numbers behind the headline
The 99% figure deserves a careful read. The test set — 3,020 phishing attacks, assembled and scored by AegisAI itself — also ran through two comparison systems that the company described only by category: a "leading closed-source general-purpose model" caught 81%, and a "widely used open-weight model" caught 80%. Neither was named, so outside researchers cannot reproduce the test. The published benchmark says nothing about false-positive rates, breakdowns by attack type, or how the sample was sourced — the operational details an enterprise security team would demand.
None of that invalidates the claim; it frames it correctly as a vendor's opening bid rather than an independently verified result. Founders launching in this space should note the playbook anyway: ship with a benchmark, be transparent about its limits, and let buyers run their own red-team samples — standard in security procurement.
Who is building it
AegisAI was founded in 2025 by Cy Khormaee and Ryan Luo, both former Google security leads whose prior work touched reCAPTCHA, Safe Browsing, and Web Risk — systems that collectively screen billions of users from phishing and malicious downloads. The company emerged from stealth in September 2025 with a $13 million seed co-led by Accel and Foundation Capital, then closed a $36 million Series A led by Battery Ventures in July 2026, bringing total disclosed funding to $49 million. Khormaee serves as CEO.
Proteus is not the company's only bet. Its platform also includes Vanguard, a threat-hunting agent pitched as an autonomous system that hunts threats broadly across an organization's environment — extending the investigative-reasoning approach beyond the inbox. The two-product structure (a narrow, benchmark-friendly flagship plus a broader agentic platform) mirrors a pattern spreading across AI security startups.
Where Proteus fits in a crowded market
Email security is far from a greenfield. Abnormal Security has built a large enterprise business on behavioral analysis; Sublime Security combines machine learning with message inspection; Material Security focuses on post-delivery mailbox protection. And the incumbents — Microsoft Defender for Office 365, Proofpoint, and Mimecast — protect the overwhelming majority of corporate inboxes, each with machine-learning detection layers added over the past several years.
AegisAI's argument is not that it is first to apply AI to email. It is that Proteus is the first model trained from the ground up as a foundational system for this single task, rather than a general-purpose model fine-tuned or prompted for security use. The broader AI security industry is splitting along exactly this line: teams that repurpose frontier models versus teams that train narrow, purpose-built ones.
Why it matters for founders and investors
Three signals for anyone building or backing security startups:
- Specialized models are fundable. AegisAI raised $49 million in roughly a year on the premise that one task deserves one model. That thesis is replicating across adjacent categories — AI agent security, runtime guardrails, threat hunting — each attracting capital on specialized-model grounds.
- The evaluation game is the marketing game. Proteus launched with a benchmark, not a parameter count, pricing sheet, or customer roster. In security, the benchmark is the launch — but buyers now expect named competitors, reproducible sets, and false-positive numbers. Founders should publish the parts competitors can't dispute.
- Incumbent budgets are the prize. Most enterprises already pay for a secure email gateway plus a behavioral-AI add-on. AegisAI's bet is that generative-AI phishing is severe enough to reopen the vendor conversation entirely — a wager the next two quarters of enterprise pilots will settle.
What to watch
Whether enterprises run side-by-side evaluations of Proteus, what pricing AegisAI reveals, and whether Vanguard's launch pulls the company into the broader agentic-SOC category. The first confirmed enterprise win will tell investors whether the 99% claim converts into revenue.
Angeled.com connects founders with investors.
Read Next
OpenAI Releases Codex Security as Open-Source CLI and SDK
Databricks Unveils Always-On Genie, an AI Agent That Stops Waiting to Be Asked
DeepMind Launches AlphaGenome 2, a DNA Model That Reads Genomes Cell by Cell
ElevenLabs Unveils 13 New Products and Upgrades at Its New York Summit
Google Launches Gemini Decision 1, a Model That Answers Instead of Writing
Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video



