AI Phishing and Email Authentication: When Wording Can't Be Trusted, the Delivery Path Must Be

AI-powered phishing has removed the oldest warning sign in email security: bad writing. For two decades, people were taught to look for misspellings, odd grammar and clumsy phrasing. Language models now produce fluent, localised, well-targeted text for almost no cost, so those cues are disappearing. The consequence for senders and defenders is that what a message says matters less, and what a message can prove about where it came from matters more. This analysis reviews the available 2025 and 2026 figures, notes the weaknesses in them, and argues that email authentication is becoming the most reliable signal left. We also look at what legitimate senders should do about it.

What the figures say

How to read them

The first figure is the loudest and the most slippery. The source describes emails that "utilised AI", which can mean anything from a model polishing a sentence to a fully automated campaign. It is not the same as "82.6% of phishing is written by AI", though the headline version often appears that way. The 37% figure is also worded differently in different places, as a share of breaches and as a share of methods. The link between rising BEC volume and AI is asserted in the roundup but not sourced.

The study that deserves most attention is the click-rate result, because it is a controlled comparison. Researchers reported that automated, AI-driven spear phishing achieved a 54% click rate, which the roundup says matches skilled human attackers. For context, the average click rate across phishing in the Verizon report was 2.7%. Targeted, personalised messages perform an order of magnitude better than generic ones, and AI makes personalisation cheap. That is the real shift: not better prose, but scale of tailoring.

Why filters and people both struggle

Traditional filters leaned on features that generated text can avoid. Repeated templates, known bad phrases and obvious language errors all fall away when each message is generated fresh. People face a parallel problem. Training that tells staff to spot errors is less useful against messages with none. Training should move towards checking the process: does the request change payment details, ask for urgency or secrecy, or arrive from an unexpected route?

Attackers also benefit from translation and localisation. A campaign that once required a native writer in each target language can now be produced in many. For an Indian business, this means convincing phishing in regional languages and in the professional English of a bank or a courier, aimed at employees and customers alike.

What AI cannot fake: authentication

Whatever a message says, the delivery path leaves technical evidence. SPF shows whether the sending server is authorised for the domain in the envelope. DKIM shows whether the message was signed by a key published by the domain. DMARC ties these to the domain in the From header that the reader sees and tells receivers what to do when they fail. A fraudster can write a perfect message, but cannot make a message pass DMARC for a domain they do not control.

This is why enforcement matters. A domain at p=none publishes a record but asks receivers to do nothing about spoofed mail. A domain at p=reject asks them to refuse it. In our analysis of DMARC adoption in 2026, EasyDMARC found a valid record on 52.1% of the top 1.8 million domains but only 22.9% at quarantine or reject. In other words, most of the domains that could be protected from direct spoofing are not. Direct spoofing is only one technique, but it is the one that authentication stops outright.

The limits of authentication

Honesty requires saying what authentication does not do:

• Lookalike domains. An attacker can register a similar domain, authenticate it perfectly and send from there. DMARC on your domain does not stop that.
• Compromised accounts. Mail sent from a hijacked mailbox passes authentication. Strong sign-in, such as passkeys, addresses this layer.
• Display-name tricks. Showing a trusted name beside an unrelated address works on readers who do not check the address.
• Misuse of trusted services. Attackers can send through legitimate platforms, so reputation signals still matter.

Authentication is necessary, not sufficient. Its value is that it removes the cheapest attack and gives providers a reliable basis for decisions.

How receivers are responding

Mailbox providers have been tightening on this basis. Gmail, Yahoo and Microsoft now require authentication from bulk senders, and in 2025 began rejecting non-compliant mail. We describe that timeline in our analysis of bulk-sender rules. They also use signals beyond content: sender reputation, complaint rates and engagement. Visual cues are another response. BIMI displays a verified logo, and a checkmark in Gmail, for senders with enforced DMARC and a certificate, as we explain in our post on BIMI in 2026. The logic is the same: reduce reliance on reading prose and increase reliance on proof.

What legitimate senders should do

Protect your domain from being spoofed

1. Move DMARC to enforcement. Use aggregate reports to find every system that sends as you, align them and then move through quarantine to reject. Our guide to aggregate reports explains the path.
2. Protect the subdomains too. Set a subdomain policy, and consider a restrictive record for domains you own but never send from.
3. Watch for lookalikes. Monitor new registrations of names close to your brand and have a process for takedown requests.

Make your real mail easy to recognise

1. Use consistent sending domains. If customers always receive payment notices from one domain, an email from a different one stands out.
2. Do not ask for sensitive actions by email. Tell customers plainly that you will never ask for passwords or card numbers, and design flows so that sign-in happens in your app or site.
3. Keep links predictable. Avoid unfamiliar shorteners and third-party tracking domains in security-critical messages. Use your own branded link domain.
4. Consider BIMI when DMARC enforcement is in place, to give recipients a visible cue.

Reduce the harm when someone is fooled

1. Use phishing-resistant sign-in. Passkeys and hardware keys cannot be typed into a fake page. See our analysis of passkeys and transactional email.
2. Verify payment changes out of band. A call to a known number should confirm any change of bank details, whatever the email says.
3. Have a response plan. Know who to call when a user reports a suspicious message, and how to revoke access quickly. For senders, our guide to blocklist removal covers recovery if your domain is abused and listed.

The senders' dilemma: you look like phishing too

There is an awkward side effect. Legitimate mail often shares features with phishing: urgent language, links, a request to act. As recipients and filters become more suspicious, honest senders pay a price. Messages from a new domain, with a shortened link and vague wording, will do worse than before. The practical answer is to look consistently like yourself: authenticated, from a known domain, with a stable name, clear purpose and an unsubscribe route. Our spam troubleshooting checklist is a good starting point for auditing how your mail appears.

Three scenarios to 2028

1. Authentication as baseline. Enforcement rises as mailbox providers push it, and direct spoofing becomes rare. Attackers shift to lookalike domains and compromised accounts. Defence moves toward domain monitoring and strong sign-in.

2. Arms race in content. Receivers deploy their own models to judge intent and context, while attackers adapt their text to evade them. Neither side wins durably, and authentication and reputation signals gain weight as the more stable inputs.

3. Agent-mediated mail. Assistants read and act on mail for people, as described in our analysis of Gmail's Gemini features and of agents as email senders. Attackers target the assistants with injected instructions as well as the people. Authenticated sender identity becomes an input to what an assistant will trust.

What we do not know

Several important questions lack reliable answers. We do not have independent measurements of how much of all phishing is machine-written, because definitions vary and detection is hard. We do not know how well mailbox providers' own models catch generated text. And we have not seen controlled data on how much BIMI or similar cues reduce successful phishing. Treat claims on these points with care, including ours.

Bottom line

If a message's wording can no longer be trusted, the delivery path must carry more of the load. Enforced DMARC, consistent sending identities, phishing-resistant sign-in and out-of-band checks for money are not new ideas. They have become the sensible response to an attacker who writes as well as you do.

Sources

• StationX: Phishing Statistics 2026 (roundup of the figures above, citing Keepnet Labs, VIPRE, Harvard Business Review, IBM, Abnormal Security, APWG and Verizon)
• EasyDMARC: 2026 DMARC Adoption and Enforcement Report