AegisAI came out of stealth just over a year ago with a simple pitch: email security needs to think like a human analyst instead of matching patterns against a blocklist. On October 8, 2026, the San Francisco startup put that pitch into a shipped product, announcing Proteus, a foundational AI model built specifically to investigate every email before it lands in an inbox. The launch, detailed in a release distributed via PR Newswire and picked up by outlets including MarkTechPost, positions Proteus as a departure from the rules-and-signatures approach that has defined email security for two decades.
The company says Proteus does not simply classify a message as spam or not-spam. Instead, it runs a structured investigation on each email, similar to what a security operations center (SOC) analyst would do by hand: checking who actually sent the message, following where embedded links lead, and reasoning about what the email is asking the recipient to do. That framing matters because it addresses the biggest weakness of legacy filters, which struggle with phishing emails that have never been seen before and contain no known-bad signatures.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What AegisAI Just Announced
The headline claim from AegisAI’s announcement is a detection number: Proteus identified more than 99% of 3,020 human-reviewed phishing attacks in the company’s internal evaluation. For comparison, AegisAI says a leading closed-source general-purpose model caught 81% of the same attack set, and a widely used open-weight model caught 80%. AegisAI has not named which general-purpose models it tested against, and the figures come from the company’s own benchmark rather than an independently audited study.
Proteus is described as a foundational model, meaning it was trained from the ground up for the email-security task rather than being a fine-tuned wrapper around an existing large language model. AegisAI has not disclosed a parameter count, pricing, or a customer count specific to Proteus. That omission is notable: companies that publish a specific accuracy figure but withhold model size, cost, and deployment scale are giving reporters and buyers a benchmark to evaluate, not a full technical specification.
Inside Proteus: How an AI Analyst Investigates Every Email
The mechanics AegisAI describes break down into three investigative steps. First, Proteus examines sender identity, looking past the display name to the underlying infrastructure, domain history, and authentication signals. Second, it follows links rather than just scanning them, attempting to understand where a URL actually redirects before a human ever clicks it. Third, and most distinctively, it reasons about intent: what action is the email trying to get the recipient to take, and does that action match the purported sender’s identity and relationship to the recipient.
That intent-reasoning step is the part that separates Proteus from most commercial email filters still in production today. Traditional secure email gateways rely heavily on reputation lists, known malware signatures, and static rule sets, which is precisely why AI-generated phishing content, novel domains, and first-contact business email compromise attempts slip through. This matters more now than it did even a year ago: attackers increasingly use generative AI to produce grammatically clean, context-aware lures that have no historical fingerprint for a signature-based filter to match against, a dynamic documented in coverage of fake AI-brand ads being used to hijack multi-factor authentication at six major AI companies earlier this year.
The Benchmark Numbers Behind the Launch
Numbers travel fast in security marketing, so it is worth laying out exactly what AegisAI has published and what it has not. The company’s benchmark used a fixed set of 3,020 phishing attacks that it says were reviewed by humans before scoring. Proteus cleared more than 99% of that set. The two comparison systems, described only by category rather than name, scored 81% and 80% respectively.
| Model Tested | Detection Rate | Category |
|---|---|---|
| Proteus (AegisAI) | >99% | Foundational model, purpose-built for email security |
| “Leading closed-source general-purpose model” | 81% | Unnamed commercial large language model |
| “Widely used open-weight model” | 80% | Unnamed open-weight large language model |
What the table does not show is a false-positive rate, a breakdown by attack type (phishing versus business email compromise versus malware delivery), or any detail on how the 3,020-email sample was sourced. Those are the numbers a buyer’s security team would actually want before signing a contract, and AegisAI has not published them alongside the headline figure.
Who Built AegisAI
AegisAI was founded in 2025 by Cy Khormaee and Ryan Luo, both former Google security leads. According to the company and investor coverage of its funding rounds, the pair’s prior work at Google touched reCAPTCHA, Safe Browsing, and Web Risk, three systems that collectively screen billions of users from phishing pages and malicious downloads across the web. That background is the core of AegisAI’s pitch to enterprise buyers: the founders spent years building abuse-detection infrastructure at Google’s scale before trying to apply the same reasoning to a dedicated email-security model.
The company is headquartered in San Francisco; earlier coverage from when it emerged from stealth in September 2025 also described offices in San Francisco and New York. Khormaee serves as chief executive. AegisAI’s public materials have not published a detailed quote from either founder specifically about the Proteus launch, so any characterization of their intent beyond the published benchmark and product description should be treated as the company’s own framing rather than a verified statement.
The Funding Trail From Stealth to Series A
AegisAI’s capital history is short but fast-moving, which is typical for AI security startups that raise a seed round, ship a product within a year, and then raise again. The company emerged from stealth in September 2025 with $13 million, co-led by Accel and Foundation Capital. Ten months later, in July 2026, it closed a $36 million Series A led by Battery Ventures, with both earlier investors returning, bringing total disclosed funding to $49 million according to SecurityWeek’s reporting on the round.
| Round | Amount | Date | Lead Investor(s) |
|---|---|---|---|
| Seed | $13 million | September 2025 | Accel and Foundation Capital (co-led) |
| Series A | $36 million | July 2026 | Battery Ventures (Accel, Foundation Capital participating) |
| Total disclosed | $49 million | As of July 2026 | — |
Foundation Capital’s investment post, titled “the overdue email security breakthrough,” frames the founders’ Google background as the differentiator that justified a second check into the same company. At the time of its 2025 stealth launch, AegisAI reported having three paying customers, including Lokker and Mesh Connect, though that figure predates Proteus and should not be read as a current customer count for the new model.
Where Proteus Fits in a Crowded Market
Email security is not a greenfield category. Abnormal Security has built a large enterprise business around behavioral analysis of identity and communication patterns to catch business email compromise. Sublime Security combines machine learning with message inspection to flag phishing and malicious links. Material Security takes a different angle entirely, focusing on what happens after a mailbox is already compromised rather than stopping the message pre-delivery. Then there are the incumbents: Microsoft Defender for Office 365, Proofpoint, and Mimecast, which collectively protect the overwhelming majority of corporate inboxes today and have each added their own machine-learning detection layers over the past several years.
AegisAI’s argument is not that it is the first company to apply AI to email. It is that Proteus is the first model trained specifically as a foundational system for this one task, rather than a general-purpose model fine-tuned or prompted for security use, or a classifier bolted onto a rules engine. Whether that distinction proves durable is a question enterprise security teams will spend the next several quarters testing in their own environments, likely by running the same kind of side-by-side comparisons that have already become common in other security tool categories, such as the ongoing endpoint detection and response comparisons between CrowdStrike, Microsoft Defender, and SentinelOne.
| Vendor | Primary Approach | Where It Competes With Proteus |
|---|---|---|
| AegisAI (Proteus) | Purpose-built foundational model, pre-delivery investigation | New entrant, benchmark-led positioning |
| Abnormal Security | Behavioral analysis of identity and communication patterns | Direct competitor in AI-native BEC detection |
| Sublime Security | ML plus message inspection rules | Overlaps on phishing and malicious-link detection |
| Material Security | Post-delivery mailbox protection and remediation | Complementary more than competitive |
| Microsoft Defender for Office 365 | Integrated Microsoft 365 security suite | Incumbent with native ecosystem advantage |
| Proofpoint / Mimecast | Established secure email gateway platforms | Legacy incumbents adding AI detection layers |
From Spam Filters to Foundational Models: A Short History of Email Security
Email security has gone through roughly three eras. The first, starting in the late 1990s, was pure pattern matching: Bayesian spam filters and blocklists that caught volume-based junk mail but did little against targeted attacks. The second era, which still dominates most enterprise deployments today, added reputation scoring, sandboxing of attachments, and behavioral heuristics layered onto secure email gateways from vendors like Proofpoint and Mimecast. That generation got meaningfully better at catching known malware families and bulk phishing campaigns, but it still largely depends on having seen something similar before.
The third era, now underway, is defined by generative AI cutting both ways. Attackers use large language models to write convincing, grammatically flawless lures at scale, eliminating one of the most reliable old tells (broken English, odd phrasing) that human recipients and legacy filters once relied on. Defenders have responded by building AI into detection, first as an add-on classifier and now, with Proteus, as a dedicated foundational model. AegisAI is betting that the arms race favors whoever builds a model specifically for the investigative reasoning task rather than repurposing a general chatbot for it.
Why Foundational Models Are Becoming Security’s New Battleground
The broader AI security industry has started splitting into two camps: companies that fine-tune or prompt-engineer existing frontier models for security tasks, and companies building narrower, purpose-trained models for a single job. Proteus falls firmly in the second camp, and AegisAI’s benchmark claim is essentially an argument for why that approach wins on this specific task. It is the same logic playing out in adjacent categories, where startups like those tracked in the AI agent security funding comparison between Zenity, HiddenLayer, and Straiker are raising capital on the premise that securing AI systems, and using AI to secure other systems, both require specialized models rather than generic ones.
The Skepticism Around a Self-Reported 99% Score
Security researchers have learned to treat vendor-published benchmarks with caution, and AegisAI’s figures invite the same scrutiny. The 3,020-email test set was assembled and scored by AegisAI itself. The two comparison models are unnamed, which makes it impossible for outside researchers to reproduce the test. And the claimed 99%-plus detection rate says nothing about the false-positive rate, which in email security is often the more operationally painful metric: a filter that blocks 1% of legitimate business mail at enterprise scale can generate thousands of help-desk tickets a month.
None of this means the benchmark is wrong. It means the number should be read as a vendor’s opening claim rather than an independently verified result, the same caveat that applies to nearly every detection-rate figure published in a product launch. Enterprise buyers evaluating Proteus will likely want to run their own red-team phishing samples against it before trusting the headline figure, a step that has become standard practice given how often nation-state actors have specifically targeted mailboxes and email infrastructure, as seen in recent reporting on the FBI tying China-linked hackers to mailbox theft campaigns with a $10 million reward now attached to the case.
Vanguard and AegisAI’s Bigger Agentic Security Bet
Proteus is not AegisAI’s only product. The company’s platform also includes Vanguard, described in funding coverage as a threat-hunting agent that AegisAI planned to push toward general availability using proceeds from its Series A. Where Proteus investigates individual emails before delivery, Vanguard is pitched as an autonomous agent that hunts for threats more broadly across an organization’s environment, extending the same investigative-reasoning approach beyond the inbox.
That two-product structure mirrors a pattern showing up across the AI security startup landscape: a narrow, benchmark-friendly flagship product (in this case, Proteus) paired with a broader agentic platform play (Vanguard) aimed at the same budget line. It is also why credential and access security for AI agents themselves has become its own urgent sub-category, covered in detail in guidance on how to secure AI agent credentials through an MCP gateway setup. If Vanguard operates with broad autonomous access across a customer’s systems, the credentials and permissions it holds become a new attack surface in their own right, not just a defensive tool.
What This Means for Enterprise Security Budgets
For CISOs, the Proteus launch adds one more line item to an already crowded email-security renewal conversation. Most enterprises already pay for a secure email gateway from an incumbent like Proofpoint, Mimecast, or Microsoft, often layered with a behavioral-AI add-on from a company like Abnormal Security. A foundational model from a 2025-founded startup with $49 million in total funding is a meaningful new option, but it is also an unproven one at the scale of a Fortune 500 mail flow, where incumbents process billions of messages a day across years of tuned detection logic.
The more interesting budget signal is what AegisAI’s rapid raise-to-launch timeline says about investor appetite in this category. Going from a $13 million seed to a $36 million Series A in ten months, then shipping a flagship model within about three months of closing that round, suggests investors are pricing in urgency around AI-native email threats rather than waiting for a slower, more traditional product-market-fit cycle. That urgency is compounded by a steady drumbeat of nation-state activity tied to email and mailbox compromise, including the kind of AI-agent-linked attacks South Korean banks have faced, detailed in reporting on CrowdStrike’s analysis tying an AI agent to China-linked bank intrusions earlier this year.
Five Predictions for AI-Native Email Defense
Based on the trajectory of AegisAI’s funding, the shape of its benchmark claim, and the broader direction of the email security market, a few outcomes look likely over the next 12 to 18 months:
- Independent third-party testing of Proteus will emerge within six to twelve months, either from a security research firm or a rival vendor seeking to rebut the 99% figure, and the real-world number will likely land lower once false positives and a broader attack-type mix are factored in.
- Incumbents including Microsoft, Proofpoint, and Mimecast will respond by marketing their own models as “foundational” or “purpose-built,” regardless of whether the underlying architecture actually changes, because the term has become a competitive buzzword AegisAI’s launch just validated.
- AegisAI will raise a Series B within 12 to 18 months if Proteus signs even a handful of named enterprise logos, following the same raise-then-prove-it cadence it used between its seed and Series A rounds.
- Vanguard, not Proteus, will become AegisAI’s larger long-term revenue driver, since broad agentic threat-hunting typically commands higher per-seat enterprise pricing than a single-channel email filter.
- At least one competitor in the AI agent security space will cite the Proteus benchmark dispute as a cautionary example in its own marketing, using it to argue for third-party-audited detection claims as a selling point against vendors that publish only self-reported numbers.
None of these are certainties, and AegisAI could just as easily prove the skeptics wrong by publishing a fuller, independently reviewed benchmark in the coming months. But the pattern of benchmark-led launches followed by independent scrutiny has repeated often enough across the AI security sector that it is a reasonable baseline expectation here too.
Frequently Asked Questions
What is AegisAI Proteus?
Proteus is a foundational AI model built by AegisAI specifically for email security. Rather than classifying messages using static rules or signatures, it investigates each email’s sender identity, embedded links, and requested action before the message reaches the recipient’s inbox.
When did AegisAI launch Proteus?
AegisAI announced Proteus on October 8, 2026, in a release distributed via PR Newswire and covered by outlets including MarkTechPost.
What detection rate does AegisAI claim for Proteus?
AegisAI says Proteus identified more than 99% of 3,020 human-reviewed phishing attacks in its internal benchmark, compared with 81% for an unnamed leading closed-source general-purpose model and 80% for an unnamed widely used open-weight model. These figures are self-reported and have not been independently audited.
Who founded AegisAI?
AegisAI was founded in 2025 by Cy Khormaee, who serves as CEO, and Ryan Luo. Both are former Google security leads whose prior work touched reCAPTCHA, Safe Browsing, and Web Risk.
How much funding has AegisAI raised?
AegisAI has raised $49 million total: a $13 million seed round in September 2025 co-led by Accel and Foundation Capital, followed by a $36 million Series A in July 2026 led by Battery Ventures.
Is Proteus available to enterprise customers now?
AegisAI’s announcement did not disclose pricing or a specific customer count for Proteus. At the time of its 2025 stealth launch, the company reported three paying customers, including Lokker and Mesh Connect, though that figure predates the Proteus model itself.
How does Proteus compare to Microsoft Defender for Office 365 and Proofpoint?
Microsoft Defender for Office 365 and Proofpoint are established secure email gateway platforms that have layered machine-learning detection onto rules-based and reputation-based filtering built over many years. AegisAI positions Proteus as a purpose-built foundational model rather than an AI layer added to an existing gateway, though it has not published a head-to-head comparison against either incumbent.
What is Vanguard, and how does it relate to Proteus?
Vanguard is a separate AegisAI product described as a threat-hunting agent, which the company planned to move toward general availability using capital from its Series A. While Proteus focuses on pre-delivery email investigation, Vanguard is positioned as a broader autonomous agent for hunting threats across an organization’s wider environment.
