AI Phishing: What Changed and How to Protect Your Company

For years, the advice against phishing was to look at the details: spelling mistakes, clumsy translations, a generic greeting. That advice is becoming obsolete. In December 2024, the FBI warned in a public service announcement that generative AI reduces the time and effort criminals need to deceive their targets, and that they use it for translation to avoid grammar and spelling errors.

For a small business, the real risk is not a strange email. It is a correct email, in good language, asking for something that looks like part of the job: pay an invoice, update a supplier's bank account, approve an urgent transfer because the manager asked for it.

This article summarizes what changed and which controls actually work, based on guidance from the FBI, CISA, NIST, and the UK's NCSC, and closes with the channels for reporting in Chile.

What changed with generative AI

The same FBI announcement describes how generative AI is used in fraud:

• Text: believable messages, error-free and at scale, plus fake profiles and fraudulent websites.
• Audio: short clips with a cloned voice to ask for money in a fake emergency.
• Video: real-time video calls with alleged company executives or authority figures.

For companies, the core fraud is business email compromise (BEC). The 2025 IC3 annual report, from the FBI's Internet Crime Complaint Center, defines it as a scam targeting those who work with suppliers or regularly make wire transfers, carried out by compromising email accounts, phone numbers, and virtual meeting applications. According to that report:

• BEC accounted for 24,768 complaints and US$3.046 billion in losses in 2025, the second-costliest crime type after investment fraud.
• Phishing and spoofing was the most frequently reported crime type, with 191,561 complaints.
• 22,364 complaints referenced artificial intelligence, with adjusted losses of US$893 million.
• Businesses reported more than US$30 million in losses to BEC involving AI: emails that mimic the CEO and cloned voices requesting wire transfers.

Two caveats. These are complaints filed in the United States, not a measurement of Chile, and we found no equivalent Chilean statistic with a public methodology. Also, IC3 flags a case as AI-related only when the complaint mentions AI, and, discussing investment fraud, the report itself warns that many victims do not realize the extent AI may be involved.

Why "look for spelling mistakes" no longer works

The UK's NCSC puts it plainly: spotting all phishing emails is hard, and training based on standard warning signs helps users spot some, but cannot teach everyone to spot them all. It adds that some attacks will always get through.

With generative AI, an error-free email in your manager's tone has nothing to spot at a glance. Defense has to move to two places: technical controls that do not depend on the human eye, and processes that force verification before paying or granting access.

The controls that actually work

1. Phishing-resistant MFA

Not every second factor protects equally. The CISA fact sheet (October 2022) explains that a fake site can ask for the password and the six-digit code from the authenticator app, then use them on the real site. SMS or voice codes are also vulnerable to SIM swap; CISA treats them as a last resort and calls phishing-resistant MFA the gold standard.

NIST SP 800-63B-4 (July 2025) is more precise: authenticators that require typing a code, such as OTPs, are not phishing-resistant, because the code is not bound to the session. According to CISA, the only widely available phishing-resistant option is FIDO/WebAuthn: physical security keys or authenticators built into the computer or phone. Passkeys, according to the FIDO Alliance, are credentials based on those same standards.

What to do:

• Turn on FIDO2 or passkeys first for email, remote access, and online banking.
• Start with the highest-value accounts. CISA suggests a first phase with system administrators and the help desk; in a small business, add the people who approve payments.
• Where that is not yet possible, use an authenticator app with codes or with number matching, which CISA considers the best options for small and medium-sized businesses in the meantime.
• Keep SMS only as a temporary solution.

2. SPF, DKIM, and DMARC with a reject policy

These DNS records let whoever receives an email check whether it came from a server authorized by your domain. The phishing guidance from CISA, NSA, FBI, and MS-ISAC (October 2023) calls for setting DMARC to "reject" for the email you send: that way, emails spoofing your domain are rejected before they reach your customers and suppliers.

DMARC was updated in May 2026: RFC 9989 replaced RFC 7489. Records still start with v=DMARC1. The RFC recommends starting at p=none to receive reports and discover every service that sends email on your behalf. For domains whose users post to mailing lists, it asks to keep p=none for at least a month, then p=quarantine for a similar period, before moving to p=reject. If you publish p=reject, it requires DKIM signing rather than relying on SPF alone.

_dmarc.yourcompany.com.  TXT  "v=DMARC1; p=none; rua=mailto:dmarc@yourcompany.com"
_dmarc.yourcompany.com.  TXT  "v=DMARC1; p=quarantine; rua=mailto:dmarc@yourcompany.com"
_dmarc.yourcompany.com.  TXT  "v=DMARC1; p=reject; rua=mailto:dmarc@yourcompany.com"

One limit: DMARC protects your exact domain. It does not stop a lookalike domain (yourcompany-cl.com) or email sent from a supplier's real, compromised account.

3. Out-of-band verification for payments and account changes

The NCSC recommends verifying every important email request through a second type of communication. The FBI announcement on BEC (September 2024) asks the same for any request to change account information. Turn it into a written rule:

• No change to a supplier's bank account is applied without calling a number that was already in your records, never the one in the email or on the invoice.
• No transfer requested by email, chat, audio, or video call is executed without confirmation through a channel different from the request.
• Payments above an amount you define require a second approval.
• The rule applies even when the CEO asks for an exception. If urgency allows skipping the rule, the attacker will manufacture urgency.

4. A code word for calls and video calls

The FBI recommends that families agree on a secret word or phrase to verify identities, and that people hang up and call back a known number. It works the same way in a company: a code word known to finance and management, never written in email or chat, and requested for any payment instruction by phone or video. If the voice sounds exactly like your boss but does not know the word, nothing gets paid.

5. Fast, blame-free reporting

The NCSC calls for a clear, simple, and quick reporting process, with feedback to whoever reports, and warns that people who fear reprisals will not report their mistakes. Turn on the report button in your email platform and define who gets notified if someone has already clicked.

6. Procedure-based training

Training has to teach procedures: which requests are always verified, through which channel, who to report to, and what to do after a click. The NCSC suggests measuring how many people reported a suspicious email, not just how many clicked.

If someone fell for it: the first hours

The IC3 report stresses that, with a fraudulent transfer, time is of the essence. In 2025, its Recovery Asset Team acted on 3,900 incidents and froze US$679 million of the US$1.164 billion in attempted theft. The mechanism is a US one, but the lesson holds: call the bank before anyone else.

1. Call your bank immediately and ask to stop or recall the transfer, with the transaction number, amount, destination account, and time.
2. Lock down access: change the affected account's password, sign out all its sessions, and check for new forwarding rules or newly registered MFA methods.
3. Preserve evidence: the email with its headers, messages, audio, receipts, and a timeline.
4. Alert the real counterpart through a known channel: the impersonated supplier, the affected customer, or your own team.
5. Report and file a complaint through the channels below.

Where to report in Chile

• CSIRT Nacional (ANCI, Chile's national cybersecurity agency): csirt.gob.cl/reportar takes reports from individuals, from essential services and operators of vital importance, and from other legal entities. It also answers on phone line 1510. If your organization is regulated under Law 21.663 and the incident has a significant effect, the CSIRT notification page says to report it through the ANCI Portal.
• Criminal complaint: according to the PDI (Chile's investigative police), it is filed in person at the nearest PDI, Carabineros, or Fiscalía (public prosecutor) office, with your national ID card, explaining the facts and providing the evidence.

Checklist

Sources

• FBI IC3 · Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (Dec. 3, 2024).
• FBI IC3 · Business Email Compromise: The $55 Billion Scam (Sept. 11, 2024).
• FBI IC3 · Internet Crime Report 2025.
• CISA · Implementing Phishing-Resistant MFA (October 2022).
• CISA, NSA, FBI, and MS-ISAC · Phishing Guidance (October 2023).
• NIST · SP 800-63B-4 (July 2025).
• NCSC · Phishing attacks: defending your organisation.
• FIDO Alliance · Passkeys.
• IETF · RFC 9989 (DMARC), RFC 7208 (SPF), and RFC 6376 (DKIM).
• CSIRT Nacional · Reportar and Cómo notificar.
• PDI · Proceso de denuncia.

At Astrum, cybersecurity is our core line of work. If you want to review your controls against phishing and email fraud, let's talk.