Anti-phishing
AI spear phishing: how to defend against it
Generative AI makes targeted phishing cheap and fluent. The email filters, sign-in methods and payment rules that still stop it.
Key takeaways
- The FBI says criminals use AI-generated text to make scams more believable.
- Phishing and spoofing drew more FBI complaints than any other crime type in 2025.
- Filters that learn contact patterns help, but a hijacked real account can pass them.
- Phishing-resistant sign-in stops stolen passwords and codes from working.
- Payment changes need a callback on a known number, whatever the email says.
AI spear phishing uses generative AI to write targeted, fluent messages that impersonate a colleague, supplier or executive. Defend in layers: authenticate your own domain with DMARC, turn on impersonation filtering, move staff to phishing-resistant sign-in, verify every payment change on a known channel, and make reporting easy. No single filter catches every message.
What is AI-driven spear phishing?
Spear phishing is a message written for one person or a small group, usually posing as someone they already trust: a leader, a supplier, an IT desk. AI changes the cost, not the idea. The FBI warns that criminals use AI-generated text "to appear believable to a reader", and that it lets them reach "a wider audience with believable content".[1]
Definition
AI spear phishing is a targeted phishing message drafted or personalized with generative AI, often using public details about the target, their role and their contacts, so it reads like a genuine request from someone they know.
The old tells were spelling mistakes and odd phrasing. Those are gone. What remains is the request itself: pay this, change that, sign in here.
How big is the problem?
Phishing and spoofing produced 191,561 complaints to the FBI's Internet Crime Complaint Center in 2025, more than any other crime type it lists.[2] Business email compromise, where the target is a payment, produced 24,768 complaints and reported losses of $3,046,598,558.[2]
AI is a growing part of that. The same report says chat generators can "quickly create official-sounding emails mimicking a company's CEO", and that businesses reported losses over $30 million to BEC scams involving AI in 2025.[2] The report is clear that not all BEC tactics are AI-enabled, and these are reported losses only.[2]
Why do older filters miss these emails?
Many spear phishing emails carry no malware and no known bad link. They may come from a freshly registered lookalike domain with valid authentication records, or from a real mailbox that an attacker has taken over. Microsoft's own documentation notes that impersonation "can pass email authentication checks (SPF, DKIM, and DMARC)" when the attacker registers a lookalike domain and publishes valid records.[4]
So signature checks and blocklists are necessary but not enough. The question a filter now has to ask is whether the sender, the relationship and the request make sense together.
What can AI-based email defenses catch?
Modern email platforms add three kinds of checks on top of authentication:
- Impersonation checks. In Microsoft Defender for Office 365, user and domain impersonation protection flag senders that look like named people or domains you choose to protect.[4]
- Contact history. Microsoft's mailbox intelligence "uses artificial intelligence (AI) to determine user email patterns with their frequent contacts".[4]
- First contact warnings. A safety tip tells recipients when they "don't often get email from" a sender.[4]
These help, and they have limits worth knowing. Microsoft notes that mailbox intelligence protection "doesn't work if the sender and recipient previously communicated via email".[4] A request from a real supplier's compromised account can therefore look normal. Raising detection thresholds also raises false positives, as the same page warns.[4] Treat any vendor claim about catch rates as something to test on your own mail.
Who can send mail in your name today?
A free growth diagnostic checks your email authentication (SPF, DKIM, DMARC), lookalike domains, domain renewals and ownership, and obvious fake profiles or pages, and names what needs a specialist firm. Your top 3 findings and a 30-day fix plan in writing within 2 business days.
Which controls stop the loss?
Controls that do not depend on anyone spotting the fake:
- DMARC at enforcement on your own domains. A "reject" policy tells receivers the domain owner considers failing mail "a clear indication that the use of the domain name is not valid".[5] It stops direct spoofing of your domain, though not lookalikes. Our guide to corporate DNS security covers the setup.
- Phishing-resistant sign-in. NIST defines phishing resistance as the ability of the authentication protocol to prevent disclosure of secrets to an impostor "without relying on the vigilance of the claimant".[6] Authenticators that involve typing in a code "SHALL NOT be considered phishing-resistant".[6] Passkeys and security keys built on WebAuthn qualify.[6]
- Out-of-band verification for money and access. The UK NCSC recommends that important email requests are verified "using a second type of communication".[3] For payment changes, that means a callback to a number already on file.
- Lookalike domain monitoring. Watch for new domains that imitate yours, as described in our typosquatting audit guide.
How should training change?
Move it from spotting to verifying. The NCSC states that "spotting all phishing emails is hard, and spear phishing attacks are even harder to detect".[3] Teach people which requests always need a second channel, whoever appears to send them.
Reporting matters as much. The NCSC asks organizations to "have an effective process for users to report phishing" and warns that people who fear reprisals will not report mistakes promptly.[3] A fast, blame-free report from the first recipient can protect everyone else on the list.
What about voice and video?
The same tools now clone voices. The FBI says criminals use AI-generated audio to impersonate public figures or personal relations to elicit payments.[1] An email followed by a confirming phone call is no longer proof. Our guide to voice deepfakes in spear phishing covers that channel, and executive impersonation covers the payment controls in detail.
Who handles a successful attack?
If credentials have been entered or money has moved, contact your bank at once and bring in an incident response firm or counsel that you contract with directly. In the US, report to IC3. Those are specialist services; prevention and detection are the parts a security program should own day to day.
Where should a team start?
- Check SPF, DKIM and DMARC on every domain, and plan the move to reject.
- Turn on impersonation protection for your leaders, finance staff and key suppliers.
- Replace code-based MFA with passkeys or security keys for high-risk roles first.
- Write a payment-change policy with callbacks, and train finance on it.
- Add a one-click report button and thank people who use it.
Our anti-phishing service reviews the public side of these controls as part of our digital asset security work.
Sources
- Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud
- 2025 IC3 Annual Report
- Phishing attacks: defending your organization
- Anti-phishing policies in Microsoft 365
- RFC 9989: Domain-Based Message Authentication, Reporting, and Conformance (DMARC)
- NIST SP 800-63B-4: Digital Identity Guidelines, Authentication and Authenticator Management
Written under our Editorial and AI Policy. Sources checked 19 Oct 2026. Reviewed by DGo Digital on 19 Oct 2026.
FAQ
Questions about AI phishing
Sometimes. Email platforms combine sender authentication, lookalike-domain checks and machine learning on contact history, and these catch many attempts. Fluent writing removes the old spelling clues, though, and a message from a real but compromised account can look normal. That is why process controls sit behind the filters.