Saltar al contenido principal

19 nov 2026

Quarterly Abnormal Detection Enhancements

A quarterly summary of significant detection enhancements delivered by the Abnormal engineering team.

Attacks sent from a real vendor account, or from someone posing as a coworker, look like routine business email. Other attacks hide the lure inside a file or behind a legitimate-looking link, or bury it under a flood of real email. This quarter, we shipped upgrades across vendor compromise and impersonation, lateral phishing, attachments and links, and spam bombs.

Broader Detection of Vendor Email Compromise

When a vendor's mailbox is compromised, the attacker's email passes SPF, DKIM, and DMARC because it is sent from the vendor's real account. A new vendor AI agent investigates suspicious vendor emails in context. It compares each one with that vendor's past emails to the customer and inspects its links and attachments. When it confirms a takeover with high confidence, it opens a vendor case and remediates the email automatically. Weekly vendor compromise cases have increased about 350% since the agent launched, averaging about 133 high-confidence verdicts a week.1

Faster Containment of Lateral Phishing and Account Takeover

Once an account is taken over, the attacker uses it to phish coworkers who already trust that account. A new lateral AI agent reviews suspicious internal emails by reading the thread and weighing the sender's normal relationships and account takeover status. High-confidence detections are remediated automatically, so verdicts and remediation arrive sooner. Since going live in September, the lateral AI agent has given verdicts on an average of about 2,100 messages a day. Lateral phishing detections confirmed by the agent also also open an account takeover case on the sending account, and detectors are better tuned to how lateral phishing spreads in corporate and education organizations.

AI Triage for User-Reported Emails

When employees report a suspicious email, AI Security Mailbox now triages it with a new AI agent classifier.

New Detectors for Calendar Invites, QR Codes, Callback Scams, and More

A new detector catches phishing hidden inside calendar invite files, including fake invites the rest of detection missed. A new callback scam model catches emails and attachments that push the recipient to call an attacker’s phone number.

New signals also expose text designed to fool readers and AI models, such as punycode lookalike subdomains and invisible Unicode characters that may be used to hide prompt-injection instructions.

A new detection layer also provides added protection against emails that impersonate a real coworker, such as fake file-sharing notifications and messages from lookalike personal email accounts.

Expanded Coverage of Threats in Attachments and Links

Fake invoices and callback-phishing notices often arrive as attachments, where the lure sits inside the file instead of the email body. Attachment text detection now reads more file types, including HTML, EML, calendar invites, and Office documents, leading to approximately 349 additional attacks detected per week over a four-week period ending October 11. Visual attachment analysis looks at how an attachment actually renders, scanning QR codes and links inside attachments to catch attacks that text-based models and logo rules miss.

Stronger Spam Bomb Protection 

Spam bombs flood a mailbox with real newsletters and sign-up confirmations, burying the emails that matter. Because most of that flood looks like ordinary graymail, it’s hard to separate from legitimate mail. Abnormal now remediates those messages even when they look like graymail, and new enhancements have made remediation more consistent and reliable. Detection now also covers lower-volume attacks, and protection extends to government customers.

Each of these enhancements builds on the same foundation: behavioral AI that understands what normal looks like for every person and vendor, so more attacks are caught and remediated automatically.


1 Based on internal Abnormal production data for US-hosted customers collected from August 1 – October 31, 2026. Weeks are complete Monday-start weeks unless noted.

Protéjase contra las amenazas de correo electrónico en constante evolución

Descubra cómo la IA conductual detecta ataques que las defensas tradicionales no logran identificar.