The best email security for AI-generated spear phishing judges the request against the sender's history, not the wording. IC3 logged more than 22,000 AI-related complaints in 2025¹. AegisAI investigates every email for what it is trying to do²². The criteria below hold on Microsoft 365, Google Workspace or both, and they name no vendor.

What you can do by the end

What is AI-generated spear phishing?

AI-generated spear phishing is a targeted email written or researched with a language model, aimed at one person or one role. The FBI says criminals use AI-generated text to appear believable to a reader in spear phishing and other fraud schemes⁵. Its 2025 report adds that chat generators can quickly create official-sounding emails that mimic a company's CEO, with phishing links or directions to wire funds³.

Businesses reported losses over $30 million to BEC scams involving AI in 2025².

How do attackers use LLMs to craft hyper-personalized spear phishing?

In one study with 101 participants, fully AI-automated phishing emails reached a 54% click-through rate, on par with human experts and against 12% for the control group¹⁷. The information the AI gathered about each target was accurate and useful in 88% of cases¹⁸.

Why do legacy filters that catch traditional phishing fail on AI-generated phishing?

Legacy filters look for known bad links and known bad senders, and they also lean on broken language. Criminals use generative AI for translation to limit grammatical or spelling errors⁶.

An attacker can register a domain, set up SPF for it and send with a From address in a different domain, because SPF validates only the MAIL FROM domain¹¹. A lookalike domain with valid DNS records can pass SPF, DKIM and DMARC⁷.

Check what it reads when there is no link or file

The FBI describes CEO-style emails that carry only directions to wire funds.

Ask what the product compares the message with. Microsoft 365 itself adds sender and recipient history plus behavioral analysis to its authentication checks¹², so ask what a new tool reads beyond that.

Question to ask a vendor: when a known name asks for a wire transfer in plain text, with no link or file, what does your product compare that request with?

Ask how it spots a lookalike or impersonated sender

Built-in protection on every cloud mailbox does not include impersonation protection; Defender for Office 365 adds user and domain impersonation protection⁸. Each anti-phishing policy protects at most 350 named users, and by default none are configured¹⁰.

Mailbox intelligence uses AI to learn user email patterns with frequent contacts⁹. Ask whether a tool does the same for every mailbox, without a list to maintain.

Question to ask a vendor: do you flag a first message from a domain one letter away from a regular contact's domain, for every user instead of a named list?

Test it on a known sender whose account is compromised

Run this test on your own mail during a trial. Take a past thread with a supplier, then ask the vendor how its product scores a reply in that thread that changes payment details. The FBI's own advice for these requests is to verify changes in account information through a secondary channel or two-factor authentication¹⁹.

Question to ask a vendor: show me the verdict on a reply inside a real thread from a known sender that asks to change where money goes.

Ask how fast it removes a message after delivery

Zero-hour auto purge in Microsoft 365 searches only the last 48 hours of delivered email, and users are not notified when it moves a message¹³. A link that was clean on delivery and weaponized later can still phish a user when rewriting is off and the client does not support Safe Links¹⁴.

Ask for the median time from arrival to verdict and from verdict to removal, measured on your mail. Ask whether removal covers forwards and every mailbox that received a copy.

Question to ask a vendor: what is your median time from arrival to removal, and does a removal reach every mailbox the message reached?

Check that payment and sign-in steps hold when detection misses

For sign-in, a Conditional Access policy in Microsoft Entra can require the built-in Phishing-resistant MFA strength, which allows a FIDO2 security key¹⁵. For payment, verify every change of account details on a second channel.

For your own domain, publish DMARC with p=reject, so receivers reject mail that fails DMARC in your name¹⁶.

Question to ask a vendor: when your product flags a payment request, who sees the flag, and does it reach the approver before the payment runs?

Which email security solutions stop AI-generated attacks in 2026?

The approaches that stop AI-generated attacks read the request and the sender relationship, and they remove mail after delivery.

How to evaluate email security for AI phishing protection

Evaluate it on your own mail with a yes or no checklist, and ask for evidence on each line.

How AegisAI handles this

AegisAI catches fake executives, vendors and payment requests, even with no link or file to scan²³. When authentication passes, it does not rely on it; the verdict rests on what this sender has sent you before and on the request itself²⁴. It opens every link, file and QR code in a sandbox, follows it to the end, and judges what it finds there²⁰. It reads mail through the API, reaches a verdict in 1.9 seconds at the median, and removes a convicted message within seconds²¹. Every verdict comes with its reason in plain language²⁵, and you can see it on your own tenant.

How to measure it in 30 days

Track three numbers from the first day of the trial.

Questions people ask

Sources