# AI phishing emails: what has changed and what still stops them

URL: https://iblockspam.com/blog/ai-phishing-emails
Last updated: 2026-10-09
Published: 2026-10-09
Topic: Phishing & threats
Author: iBlockSpam Team

> How generative AI changes phishing, from flawless grammar to cloned voices, according to the FBI and NCSC, and which defenses still stop AI-written attacks.

## Short answer

Generative AI lets criminals write fluent, personalized phishing emails in any language and back them up with cloned voices or video. Spelling mistakes are no longer a reliable warning sign. What still works: DMARC on your domain, filtering on sender and link signals, phishing-resistant MFA, and verifying payment or access requests through a channel you already trust.

Phishing used to be easy to spot: odd grammar, a generic greeting, a clumsy logo. Generative AI removes most of those tells, so the [signs of a phishing email](https://iblockspam.com/blog/how-to-spot-a-phishing-email) that still work are about the sender and the request, not the writing. This guide summarizes what the FBI, the UK NCSC and CISA have actually said about AI-assisted phishing, without the hype, and which defenses still work when the writing is perfect.

## How does AI change phishing emails?

Security agencies agree on the direction, if not on dramatic numbers.

- **Fewer obvious mistakes.** The FBI says AI tools "can correct for human errors that might otherwise serve as warning signs of fraud," and that criminals use AI translation "to limit grammatical or spelling errors for foreign criminal actors targeting US victims" ([FBI IC3 PSA, December 2024](https://www.ic3.gov/PSA/2024/PSA241203)).
- **Less effort per victim.** The same PSA says generative AI "reduces the time and effort criminals must expend to deceive their targets."
- **Better research and personalization.** The UK NCSC judged that AI "provides capability uplift in reconnaissance and social engineering," and enables lures "without the translation, spelling and grammatical mistakes that often reveal phishing" ([NCSC, January 2024](https://www.ncsc.gov.uk/report/impact-of-ai-on-cyber-threat)).
- **More volume.** The NCSC's headline judgment was that AI "will almost certainly increase the volume and heighten the impact of cyber attacks" over the following two years.
- **Harder for everyone to spot.** The NCSC expected AI to make it difficult for everyone, "regardless of their level of cyber security understanding," to identify phishing, spoofing or social engineering.

What AI doesn't change is the mechanics. The message still has to come from some domain, travel through some mail server, and ask the recipient to do something: click, sign in, open a file, pay or reply.

## What does AI-assisted phishing look like in practice?

| Old warning sign | What AI changes | What still gives it away |
|---|---|---|
| Spelling and grammar errors | Fluent, natural writing in any language | The request itself: payment change, login, secrecy |
| Generic "Dear customer" | Names, job titles, real projects and colleagues | A sender address that doesn't match the claimed person |
| Odd tone for the sender | Writing style mimicked from public or stolen text | Unusual timing, urgency, or a new channel |
| One clumsy email | Back-and-forth conversation that adapts to replies | Refusal to verify by phone or in person |
| Email only | Follow-up voice notes, calls or video | A callback on a known number fails to confirm it |

### Voice and video follow-ups

The FBI's December 2024 PSA describes criminals using AI to "generate short audio clips containing a loved one's voice" and to "generate videos for real time video chats with alleged company executives." In May 2025, the FBI warned that criminals were sending "AI-generated voice messages" posing as senior US officials, combined with text messages that led to fake login pages ([FBI IC3 PSA, May 2025](https://www.ic3.gov/PSA/2025/PSA250515)).

For businesses, the risk is a two-step attack: an email asks for a payment, then a convincing voice call "confirms" it. The FBI's [2025 IC3 Annual Report](https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf) notes that "chat generators can quickly create official-sounding emails mimicking a company's CEO" and that "voice cloning can also be used to request wire payment."

## What do the official numbers say?

The 2025 IC3 Annual Report has a section on artificial intelligence. Its figures are based on complaints where the victim's report contains a reference to AI:

| Measure (2025, IC3 complaints) | Figure |
|---|---|
| Complaints with an AI reference | 22,364 |
| Adjusted losses in those complaints | $893,346,472 |
| BEC losses involving AI reported by businesses | Over $30 million |

Two cautions. First, these count only what victims reported and recognized as AI-related. Second, the same report says "there are multiple BEC tactics, and not all are AI-enabled." Total reported BEC losses in 2025 were far larger, which is covered in our [BEC prevention guide](https://iblockspam.com/blog/business-email-compromise-prevention). AI makes existing scams more convincing. It hasn't replaced them.

## Which defenses still work against AI-written phishing?

The good news: most effective controls never relied on spotting bad grammar.

### 1. Authentication on your own domain

An AI can write a perfect email "from" your CFO, but it can't make that email pass DMARC for your domain. CISA's joint guidance tells organizations to "ensure DMARC is set to 'reject' for sent emails" ([CISA and partners, October 2023](https://www.cisa.gov/sites/default/files/2023-10/Phishing%20Guidance%20-%20Stopping%20the%20Attack%20Cycle%20at%20Phase%20One_508c.pdf)). Check your current policy with the [DMARC Checker](https://iblockspam.com/tools/dmarc-checker), and follow [how to stop email spoofing of your domain](https://iblockspam.com/blog/stop-email-spoofing-of-your-domain).

DMARC has limits. Microsoft notes that impersonation "can pass email authentication checks (SPF, DKIM, and DMARC) if the attacker created a lookalike domain and published valid DNS records" ([Microsoft Learn](https://learn.microsoft.com/en-us/defender-office-365/anti-phishing-policies-about)). That's why the next layer matters.

### 2. Filtering that looks beyond the words

Good filters weigh signals that fluent writing doesn't change: the sending server's reputation, authentication results, how new the domain is, whether the domain resembles yours or a partner's, and the links and attachments. Impersonation checks, external-sender tags and first-contact warnings all work regardless of how well the email is written. Our guide to [how spam filters work](https://iblockspam.com/blog/how-spam-filters-work) explains the layers, and [lookalike domain phishing](https://iblockspam.com/blog/lookalike-domain-phishing) covers the domains AI-assisted campaigns often send from.

### 3. Verification procedures

This is the control AI can't write its way past. The FBI's advice is to "verify the identity of the person calling you by hanging up the phone," then call back on a number you look up yourself. It also suggests agreeing a secret word or phrase with family members. A business version:

1. **No payment or bank-detail change on email alone.** Call back on a number from your own records.
2. **No acting on voice notes or calls that arrive with a request.** A familiar voice is not proof.
3. **Agreed verification for executives.** A known callback number, or an internal channel, for anything urgent.
4. **Two people for payments** above a threshold, and for any first payment to a new account.
5. **Help desk checks.** Verify a caller's identity before resetting passwords or MFA.

Our guide to [vendor invoice fraud](https://iblockspam.com/blog/vendor-invoice-fraud-payment-redirection) has a full payment-change procedure.

### 4. Phishing-resistant MFA

If a fluent email does get a password, MFA limits the damage. CISA says "any MFA is better than no MFA" and identifies FIDO/WebAuthn as "the only widely available phishing-resistant authentication" ([CISA](https://www.cisa.gov/MFA)). Prioritize administrators, finance staff and executives.

### 5. Training that targets the request, not the typos

CISA recommends regular training on social engineering and phishing, and stresses the importance of reporting suspicious emails. Update your examples. Show staff well-written lures, voice-note follow-ups and QR codes (see [QR code phishing](https://iblockspam.com/blog/qr-code-phishing-quishing)). Teach three habits:

- Check the real sender address, not the display name. The [Email Header Analyzer](https://iblockspam.com/tools/email-header-analyzer) shows where a message actually came from.
- Pause on any request for money, credentials, codes or secrecy.
- Report fast, even after clicking. A quick report helps the team find other recipients.

## What should staff do with a message that looks perfect but feels wrong?

Give people a short routine they can follow without deciding whether AI was involved:

1. **Stop.** Don't click, reply, scan a code or call a number from the message.
2. **Check the sender.** Expand the From address. Compare the domain letter by letter with the real one.
3. **Name the request.** Money, credentials, a code, a file or secrecy? Any of these means verify first.
4. **Verify out of band.** Use a phone number, chat channel or meeting you already trust, never contact details from the message.
5. **Report it.** Use the report button in your mail client, even if the request turned out to be genuine. False alarms cost minutes; missed fraud can cost far more.

## Checklist: AI phishing readiness

| Control | Owner | Status to aim for |
|---|---|---|
| DMARC on all your domains | IT | `p=reject` |
| Inbound filtering with impersonation and lookalike checks | IT | On, tuned for executives and suppliers |
| External-sender tags and first-contact warnings | IT | On |
| Phishing-resistant MFA | IT | Admins, finance and executives first |
| Payment-change callback procedure | Finance | Written, followed, recorded |
| Voice and video verification rule | Leadership | Agreed and communicated |
| Training with realistic, well-written examples | HR and IT | At least yearly, plus quick refreshers |
| One-click reporting | IT and staff | Available in every mail client |

## Next steps

1. Rewrite your phishing training examples so none of them rely on typos.
2. Agree a callback rule for payments and access requests, and tell everyone, including executives.
3. Check your domain's DMARC, SPF and DKIM with the free [Email Health Check](https://iblockspam.com/tools/email-health-check).
4. If someone has already clicked or replied, follow [what to do after clicking a phishing link](https://iblockspam.com/blog/what-to-do-after-clicking-phishing-link).
5. Put filtering in front of your mailboxes that judges senders, links and attachments, not just wording. iBlockSpam filters phishing, spoofing and malware at the MX for Microsoft 365, Google Workspace and cPanel; see [how it works](https://iblockspam.com/how-it-works).

If you're adding a filtering layer for this, iBlockSpam's [pricing](https://iblockspam.com/pricing) is per domain with every mailbox included, and the [portal demo](https://iblockspam.com/demo) shows what you'd see.

Browse our other [phishing and threat guides](https://iblockspam.com/blog/topic/threats).

## Frequently asked questions

### Can you tell if a phishing email was written by AI?

Not reliably, and it doesn't matter much. The FBI says AI tools can correct the errors that used to give fraud away. Judge the request instead of the writing: an unexpected payment change, a login link, secrecy or urgency. Then check the sender's real address and verify through a channel you already trust.

### Do spelling and grammar mistakes still indicate phishing?

They can, but their absence proves nothing. The UK NCSC judged in 2024 that generative AI lets attackers create lures without the translation, spelling and grammatical mistakes that often reveal phishing. Train staff to look at what the message asks for and where it really came from.

### How do criminals use voice cloning in business email fraud?

The FBI's 2025 annual report says voice cloning can be used to request a wire payment, alongside chat-generated emails that mimic a CEO. A cloned voice may follow up an email to make a payment request seem confirmed. Hang up and call back on a number you already hold before acting.

### How much money has AI-enabled fraud cost US businesses?

In its 2025 annual report, the FBI's IC3 says businesses reported losses of over $30 million to BEC scams involving AI. Across all crime types it recorded 22,364 complaints with an AI reference and adjusted losses of $893,346,472. These are reported complaints only, so the true figure is likely higher.

### Does security awareness training still work against AI phishing?

Yes, if it changes focus. Teaching people to spot typos is less useful now. Teach them to recognize risky requests, to check the real sender address, to verify through a known channel, and to report quickly. CISA's phishing guidance recommends regular training and stresses the importance of reporting.

## Sources

- [FBI IC3 PSA: Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (December 2024)](https://www.ic3.gov/PSA/2024/PSA241203)
- [FBI IC3 PSA: Senior US Officials Impersonated in Malicious Messaging Campaign (May 2025)](https://www.ic3.gov/PSA/2025/PSA250515)
- [FBI IC3: 2025 IC3 Annual Report](https://www.ic3.gov/AnnualReport/Reports/2025_IC3Report.pdf)
- [UK NCSC: The near-term impact of AI on the cyber threat (January 2024)](https://www.ncsc.gov.uk/report/impact-of-ai-on-cyber-threat)
- [CISA, NSA, FBI, MS-ISAC: Phishing Guidance: Stopping the Attack Cycle at Phase One](https://www.cisa.gov/sites/default/files/2023-10/Phishing%20Guidance%20-%20Stopping%20the%20Attack%20Cycle%20at%20Phase%20One_508c.pdf)
- [Microsoft Learn: Anti-phishing policies in Microsoft 365](https://learn.microsoft.com/en-us/defender-office-365/anti-phishing-policies-about)
- [CISA: More than a password (multifactor authentication)](https://www.cisa.gov/MFA)

## Free tools for this guide

- [Email Header Analyzer](https://iblockspam.com/tools/email-header-analyzer): Trace hops and authentication results — runs entirely in your browser.
- [DMARC Checker](https://iblockspam.com/tools/dmarc-checker): Check your policy, reporting and the safe path from p=none to reject.
- [Email Health Check](https://iblockspam.com/tools/email-health-check): Grade your domain's MX, SPF, DKIM, DMARC and transport security in one scan.

---

Source: iBlockSpam, https://iblockspam.com/blog/ai-phishing-emails (this is a Markdown copy; the HTML page is canonical). Quote with attribution and a link. Site index for AI assistants: https://iblockspam.com/llms.txt