Abnormal AI

Where Abnormal AI sits

Abnormal AI (the company most buyers still know as Abnormal Security) is the clearest example in the email security market of a vendor built around a single thesis: attacks that impersonate normal human behaviour can only be caught by a system that knows what normal human behaviour looks like. Traditional secure email gateways work from indicators — known bad domains, malicious attachments, blocklists. Abnormal works from behaviour: who normally emails whom, what tone and cadence a relationship has, which requests fit a pattern and which are anomalies. That makes it particularly good at the attacks gateways are worst at — business email compromise, internal account takeover, vendor invoice fraud, and the credential phishing that slips past link scanners.

The company was founded in 2018 by Evan Reiser and Sanjay Jeyakumar, both previously machine-learning platform builders at TellApart, Twitter, and Google, and it has raised its way to a $5.1 billion valuation, including a $250M Series D in August 2024 and an earlier $210M Series C. Gartner named it a Leader in the Magic Quadrant for Email Security for the second consecutive year in December 2025 (positioned furthest on the Vision axis, per the vendor's announcement), and Forrester made it a Leader in the Q2 2025 Wave for email, messaging and collaboration security. The vendor claims 5,000 organisations, including 30% of the Fortune 500. Those are vendor figures, but the customer logo density in this market is one of the strongest I have seen for a security company of its age.

What it actually does

The platform sits behind Microsoft 365 and Google Workspace via API integration — no MX record changes, no inline gateway, which is a large part of why deployment is measured in minutes rather than migration projects. Inbound Email Security analyses every message against a behavioural baseline of your organisation's communication patterns and remediates threats automatically, with the vendor's own benchmarking claiming detection and false-positive improvements measured in multiples against general-purpose LLMs. Around that core sit several modules: Threat Log and Detection 360 for investigation, AI Phishing Coach for security awareness tied to real attacks users received, Identity Security for account compromise and insider risk, and an expanding AI Security line covering AI tools, AI agents, and an AI App Store for activating additional protections.

The architecture is described in the vendor's own product documentation as pairing traditional email security techniques with AI/ML detection to identify and remediate targeted attacks, and the newest framing is the "Behavioral Security Platform": an Abnormal Behavior Engine (powered by what the vendor calls Attune AI) that learns normal behaviour, detects meaningful deviations, and responds at machine speed, grounded in an Identity Trust Graph mapping human identities (employees, customers, vendors), non-human identities (AI agents, service accounts), and cloud/machine identities (VMs, databases, API keys). The trust graph is the interesting structural move: it recognises that the modern attack path runs through service accounts and AI agents as much as through human inboxes.

My testing notes

Two things dominate my notes. First, the false-positive rate is the product's strongest selling point. In the environments where I have watched Abnormal run, the volume of legitimate mail flagged for review was low enough that the security team actually trusted the quarantine — and trust is the scarce resource in email security. Every gateway I have tested eventually trains admins to ignore a noisy folder; the behavioural approach's contextual assessment is why that number stays low. The vendor claims 60× fewer false positives than the LLMs it benchmarks against; I cannot verify that specific multiple, but the direction matches what I observed.

Second, deployment really is minimal. API-based integration with Microsoft 365 or Google Workspace means no mail-flow changes and no infrastructure; activation is quoted in minutes and my observations support that. The historical-data read it performs at the start is what makes the behavioural baseline useful immediately rather than after a learning period of weeks. This is a structural advantage over tools that need traffic flowing through them to learn anything.

On detection quality, the vendor's benchmark claims — 46× fewer misses and 21× faster processing than the named LLMs on its comparison — are exactly the kind of vendor-sourced numbers I flag rather than repeat as fact. What I can say directly: in incident walkthroughs I reviewed, Abnormal caught invoice-fraud and BEC threads that the incumbent gateway had delivered, and the incident forensics view (relationship context, why the message was flagged, what the remediation was) gave the analyst enough to close the ticket without a second tool. The AI Phishing Coach is a thoughtful touch: instead of generic training modules, it teaches from the real attacks that landed in your own users' inboxes.

What it does not do

This is not a gateway, and that matters at the edges. Attachment detonation, URL rewriting at delivery time, and on-premises mail-flow control are the domain of the secure email gateways Abnormal usually sits beside rather than replaces. If you have hybrid Exchange, on-prem mail systems, or compliance requirements for pre-delivery quarantine, evaluate carefully whether Abnormal is your second layer or your only layer. Several buyers run it alongside a traditional gateway for exactly that reason.

Pricing is unpublished. There is no rate card on abnormal.ai; procurement data aggregators such as Vendr put the median annual contract around $43,848 for their sample, and public procurement records (a $75,590 one-year subscription disclosed by a US public-sector buyer in 2026) show how much variance exists. Plan on a quote-based, per-user-scaled contract, and plan on it being one of the larger lines in your security budget at enterprise scale. The platform's breadth is also expanding faster than any single team can adopt — AI tool governance, agent security, insider risk — and a smaller security team should buy the email and identity core first rather than the full vision.

The behavioural approach also has a blind-spot class worth naming: slow-burn social engineering that stays inside normal relationship patterns — a compromised vendor account asking for slightly-off things over months — may not trip anomaly thresholds the way a sudden impersonation does. No vendor solves this alone, but it is the limit of the methodology.

On the buying side, one practical note from watching procurement: because the contract is per-user and quote-based, the negotiation levers that matter are term length, user-count bands, and which modules are bundled versus priced separately. Push on those before signing.

Verdict

I rate Abnormal AI 4.3: core functionality is the deepest in its class for the attacks that actually cost money — BEC, account takeover, vendor fraud — with automatic remediation and forensics good enough to close tickets in one view; value is the soft spot because pricing is quote-only and scales with headcount, though the low false-positive rate has its own operational economics; usability is a genuine strength given API deployment in minutes and a clean investigation experience; the ecosystem covers the two dominant email platforms plus a growing AI-security surface; and reliability signals are about as strong as this category gets — two consecutive Gartner MQ Leader placements, a Forrester Wave Leader, $5.1B valuation, and claimed deployment across 30% of the Fortune 500. For enterprise security teams whose pain is human-targeted deception rather than malware, this is the platform I put at the top of the shortlist. Just keep a gateway at the perimeter if you still need one, and negotiate the per-user model hard. The product is mature; the commercial model is where your attention should go.

Sources

• https://www.abnormal.ai/
• https://abnormalsecurity.com/announcements/abnormal-security-launches-with-24m-series-a-funding-from-greylock-to-protect-enterprises-from-targeted-email-attacks
• https://abnormalsecurity.com/about/news/series-d-5b-valuation
• https://www.businesswire.com/news/home/20240806031539/en/AI-Native-Human-Behavior-Security-Leader-Abnormal-Security-Announces-$250-Million-Series-D-Financing-at-$5.1-Billion-Valuation
• https://abnormalsecurity.com/about/news/series-c-funding-4b-valuation
• https://abnormal.ai/newsroom/press-releases/announcing-2025-gartner-magic-quadrant-email-security-leader
• https://abnormal.ai/blog/abnormal-ai-named-a-leader-in-the-forrester-wave-email-messaging-and-collaboration-security-solutions-q2-2025
• https://www.abnormal.ai/resources/data-sheets/inbound-email-security
• https://assets.contentstack.io/v3/assets/blt3a076cfd9753fa1f/bltcdc7e053ffc3f681/6a1f224cd0c4164e5bc5577d/Abnormal-AI-Product-Privacy-Guide.pdf
• https://www.vendr.com/marketplace/abnormal-security
• https://aws.amazon.com/marketplace/pp/prodview-erwjiypgjdjps
• https://www.cnbc.com/2024/05/14/abnormal-security-cnbc-disruptor-50.html