AegisAI Perspectives · 024
Enterprise email security in 2026 pairs enforced DMARC and platform filtering with detection that reads each email's intent and pulls mail after delivery.
Enterprise email security in 2026 is three layers: authenticated domains, platform filtering, plus detection that reads each email's intent and removes it after delivery. AegisAI's Q3 2026 State of the AI Attack report found phishing rose 49% this quarter1. That sample covers more than 100 million emails sent to over 1 million mailboxes5.
A message from a hijacked mailbox passes the sender's own domain checks, so the program also watches the account behind the mail.
A 2026 program puts seven controls in place, domain first.
Publish SPF and DKIM records for every domain you own, including unused ones. Microsoft states that SPF, DKIM and DMARC in DNS let Microsoft 365 protect more accurately against spoofing19, and Google gives the same advice for Gmail20.
Move the DMARC policy from p=none to p=quarantine to p=reject. RFC 7489 defines that range, from no action through altered delivery up to rejection21. Our guide to rolling out DMARC without breaking mail covers the order.
Add TLS and S/MIME on top of the three DNS records, following NIST SP 800-177 Rev. 122.
Switch on the platform's own filtering. Defender for Office 365 Plan 1 covers zero-day malware, phishing and business email compromise18.
Block the device code flow with a Conditional Access policy. Run it in report-only mode first to see who still uses the flow, then switch it on10.
Alert on new MFA methods and unusual sign-ins in the identity log. AI Dragnet attacks sent from hijacked accounts rose 38x in Q3 2026, from 0.2% to 7.7%2.
Turn on automated investigation and response, so a bad email is removed from every mailbox after delivery. Plan 2 adds investigation, hunting, response and automation.
Steps 1 to 3 stop exact-domain spoofing. Steps 4 to 7 deal with mail that passes those checks.
AegisAI reads mail through the Microsoft 365 or Google Workspace API after delivery, as the inbound email security page describes23. Every verdict comes with its reason in plain language, so the security team can check it and share it24. When one bad email lands, it is pulled from every inbox it reached, forwards included25.
Five figures from two primary sources each point at one control.
| Figure | Source | What it measures | Control it points to |
|---|---|---|---|
| 49% | Q3 2026 State of the AI Attack report | Rise in phishing this quarter | Detection that reads intent, plus removal after delivery |
| 38x (0.2% to 7.7%) | Q3 2026 State of the AI Attack report | AI Dragnet attacks sent from hijacked accounts | Account signals, since the real account passes DMARC |
| 73% | Q3 2026 State of the AI Attack report3 | Tracked organizations reached by one calendar-invite worm | Inspection of calendar invites as well as messages |
| $3,046,598,558 | FBI IC3 202513 | Business email compromise losses reported in 2025 | Payment change checks and BEC detection |
| 191,561 | FBI IC3 202514 | Phishing and spoofing complaints in 2025 | DMARC at p=reject for exact-domain spoofing |
The IC3 counted 1,008,597 complaints in 2025, with $20.877 billion in reported losses12. Business email compromise came second by loss, after investment fraud. For the wider set of breach figures and their limits, see what share of breaches start with a phishing email.
A 2026 kit checks the visitor before it shows the lure, and Wazza runs four such checks6.
Real case, Threat Watch, 2026-10-15. Wazza: A Device Code Phishing Kit That Decides Who Sees the Lure
Wazza runs every visitor through four checks before it shows an Adobe-themed device code page6. Each visit first passes a hostname check, a Cloudflare Workers beacon, a signed session token and an anti-bot gate7. The kit takes 6 steps from the landing domain to the device code page8. It treats a scanner differently from a person9. The page advises running the Conditional Access policy in report-only mode first, then switching it on.
A link scanner that follows the URL is treated differently from a person, so it never sees the lure. Detection has to judge the email itself: who sent it, what it asks the reader to do, and whether that fits the relationship. The Microsoft 365 settings that close this route are in how to protect Microsoft 365 from device code phishing.
Device code lures also reach GitHub accounts, where GhostAction's October 2026 wave targeted more than 30,000 repositories through about 1,900 compromised accounts11.
It is shifting because more attacks now arrive from real accounts, which pass the domain checks the first layer relies on. AI Dragnet attacks sent from hijacked accounts rose 38x in one quarter, from 0.2% to 7.7%. A message from a real mailbox carries that domain's own SPF and DKIM. DMARC then passes, because the domain is genuine.
The 2026 Verizon DBIR puts software vulnerabilities first, at 31% of breaches, ahead of stolen passwords17. Email controls still answer business email compromise, the second largest loss category in the 2025 IC3 report.
Yes, because the architecture decides where mail is read and what changes in mail flow. A secure email gateway sits in the mail path and receives traffic through an MX record change, as our guide on how to replace a secure email gateway without breaking mail flow explains. An API-native layer such as AegisAI reads mail through the Microsoft 365 or Google Workspace API after delivery.
An API-native layer works beside the platform's own filtering, which Defender for Office 365 provides in Plan 1 and Plan 2. It also acts on mail that already landed, so removal after delivery becomes part of the test.
LLM-generated phishing removes spelling and grammar as a test, so detection has to judge what a message asks for and who is asking. The report counted a 38x rise in AI Dragnet attacks from hijacked accounts, and one calendar-invite worm reached 73% of tracked organizations.
The threat model now includes the account behind the message, since GhostAction's October 2026 wave worked through about 1,900 compromised accounts.
Four criteria decide most evaluations: where mail is read, how a verdict is explained, how a delivered email is removed, and which account signals are used.
For a full scoring method, use the framework for evaluating email security vendors.
For the domain layer, NIST SP 800-177 Rev. 1 gives recommendations for trust in email, covering SPF, DKIM, DMARC, TLS and S/MIME. It is NIST guidance rather than a sector regulation, so confirm which rule applies to your sector with your compliance team. A DMARC record states its policy in DNS, which an auditor can read directly.
For the detection layer, keep the verdict reason and the removal record for each message. Business email compromise produced $3,046,598,558 in reported 2025 losses, and the removal record shows which mailboxes held a fraud email and when it was pulled.
Build the roadmap in the order the controls depend on each other, domain first.
Review the roadmap each quarter against the report figures above. A rise like the 49% in Q3 2026 is the signal to retest the detection layer.
Track three numbers for one month, each from a log you already have.
Compare the first week with the fourth. A falling time to remove with flat false positives means the response layer works.
It is a layered program. Domains publish DMARC at p=reject. The platform's built-in filtering checks every message. A detection layer watches accounts and removes bad mail after delivery. Phishing rose 49% in Q3 2026.
No. DMARC at p=reject stops mail that fails authentication for your domain. Mail sent from a real but hijacked account passes, and those attacks rose 38x in one quarter.
It is the base layer. Defender for Office 365 Plan 1 covers phishing plus business email compromise, and Plan 2 adds investigation and automated response. Account signals and removal after delivery sit on top.
It tricks a user into entering a code that signs the attacker in. Wazza runs four checks on each visitor before it shows the device code page. Blocking the flow with Conditional Access closes the route.
The FBI IC3 recorded $3,046,598,558 in business email compromise losses in 2025, the second largest loss category after investment fraud. It received 24,768 such complaints that year.
