AegisAI Perspectives · 121
Real-time AI adaptation stops zero-day email by judging each request against the sender history; this page gives an eight-step trial for your own mail.
See it on your own Google Workspace or Microsoft 365 tenant. No MX change.
Or get the Q3 2026 State of the AI Attack report (free, instant download)
Not ready for a demo? Check your domain (free, about 10 seconds).
Real-time AI adaptation protects against zero-day email by judging what each message asks for and who sent it, because a new lure has no signature or reputation to match16. AegisAI investigates every inbound email for what it is trying to do and removes the ones that fail from every mailbox they reached18.
The model replaces the lookup "have I seen this before?" with two checks: what the message asks for, and whether this sender usually asks for it. NIST defines a zero day attack as one that exploits a previously unknown vulnerability1. In email the unknown part is the lure itself: AI writes every lure fresh, from authenticated or compromised senders17.
The AegisAI product page states there is nothing to write or maintain as attacks change22, and how zero-day email attacks evade traditional security filters covers why signature filters miss these lures.
Write down the three moves and ask every candidate to show each one on a message from your tenant.
Short on time? Watch the explainer, or book a demo.
The first copy of a new lure arrives before any update exists: Microsoft updates spam and malware signatures in real time on a daily basis, yet users can still receive zero-day malware that was undetectable during mail flow9. Its zero-hour auto purge looks back over only the last 48 hours of delivered email.
In a study submitted on 30 November 2024, fully AI-automated spear phishing reached a 54% click-through rate, level with human experts2. The control group of arbitrary phishing reached 12%. In a dataset of more than 20,000 production phishing emails, 50.3% of AI-generated phishing got past the built-in Gmail and Microsoft filters, against 28.5% of human-written phishing24. The Q3 2026 State of the AI Attack report found phishing rose 49% in the quarter25.
Count how many of last month's reported phishing messages reached inboxes before any signature or purge acted on them.
It reads who the sender is to you, because attackers can introduce small changes in the statistical characteristics of their emails to evade machine learning detectors4. A profile-based model, built on each sender's normal pattern, delivered 30% less false positive and 25% less false negative classifications than an ML ensemble in that 2022 study.
Microsoft's Safe Links detonates URLs without a valid reputation in the background14. In one documented Safe Links scenario, a link that was clean on delivery and weaponized later went undetected, because it wasn't malicious when checked.
The NCSC says a layered approach gives you multiple opportunities to detect a phishing attack6, and that reliance on any single layer would have missed some of the attacks.
Ask each vendor which sender history its model keeps, and for how long.
Run these steps on your own tenant for one month, with the current stack left in place.
Count the phishing messages users reported last month and the messages your current stack removed after delivery.
Connect a tenant, run in monitoring mode, and compare against what your current stack delivered21. Monitoring mode comes first, and removal needs write access your admin grants and can revoke.
Microsoft 365 filters and ZAP take no action on messages it marks as phishing simulations12, so the current stack never gets a fair chance at an allowlisted test.
Count the messages the new layer flags that the current stack delivered; zero-day lures arrive unannounced, so that count is the result.
Write a payment request with no link or file from a mailbox that has never written to the recipient, and note which layer flags it.
Each verdict comes with its reason in plain language; mark it right or wrong.
Note delivery time and verdict time for every flagged message, and compare the median with the 1.9 seconds the product page states.
Log every legitimate message the new layer would have removed, by sender and by business process.
Write the three decision numbers into the trial plan before you connect anything.
Score each approach on the same rows during the trial.
| Criterion | Signature and reputation filtering | Real-time adaptation in AegisAI |
|---|---|---|
| A lure never seen before | Signatures and reputation have nothing to match | Verdict rests on the request and the sender's history |
| How it keeps up | Signature updates on a daily basis | Nothing to write or maintain as attacks change |
| Mail already in inboxes | Purge looks back 48 hours | Removed from every mailbox it reached |
| Evidence for the analyst | A verdict label | Reason in plain language |
| Filtering before delivery | Yes, it acts during mail flow | No edge. It reads mail through the API after delivery |
The last row gives in-line filtering the edge, since the API layer reads mail after delivery. Keep both layers through the trial and score all five rows.
AI agents investigate every inbound email for what it is trying to do, and remove the ones that fail from every mailbox they reached. Signatures and reputation have nothing to match. Nothing to write or maintain as attacks change. Every verdict comes with its reason in plain language, so your team can check it and share it. Runs in monitoring mode first.
Illustrative example, figures are fictional: Harbor Point Clinics (fictional) has 900 mailboxes, runs the eight steps for 30 days and records 17 user-reported phishing messages in step 1. In monitoring mode the new layer flags 41 messages the current stack delivered, 6 of them payment requests with no link, and it flags the team's own fresh lure from step 5, which the current stack delivers. Step 6 marks 39 of 41 reasons right, step 8 logs 3 false positives from one new supplier, and the team turns on removal on day 31.
Copy the structure and replace every fictional figure with one from your own trial.
Track three numbers, each from a system you already run.
Compare the three numbers with the step 1 baseline on day 30 and decide on removal.
It judges the request and the sender's history instead of matching known signatures, so a new lure is caught on first sight. A failing message is removed from every mailbox it reached.
NIST defines a zero day attack as one that exploits a previously unknown hardware, firmware or software vulnerability, and in email the new element is usually the lure, the sender or the link.
Yes, keep them: the NCSC says reliance on any single layer would have missed some of the attacks, and some phishing will always get through, so plan for incidents.
No. Microsoft 365 filters and ZAP take no action on messages the policy marks as simulations, and each entry needs a domain and a sending IP. Send trial lures as ordinary mail.
The product page states a median verdict of 1.9 seconds, with removal from every mailbox within seconds.
