Key Takeaways
- LLMs trained on corporate communication patterns can generate highly personalised phishing at scale
- Business Email Compromise (BEC) losses reported to the FBI's Internet Crime Complaint Center (IC3) topped US$3 billion in 2025
- Voice cloning has enabled AI phone phishing (vishing) that is equally convincing to email attacks
- Technical controls (SPF, DKIM, DMARC) remain essential but are insufficient — human awareness is the last line of defence
The Facts
AI-generated phishing emails, crafted using large language models, can closely mimic the tone and format of genuine business correspondence, and they lack the spelling mistakes and awkward phrasing that employees have long been taught to look for. That removes one of the easiest signals for spotting a phishing attempt.
The mechanism of advantage is personalisation at scale. A human phishing operator can craft a convincing personalised email for high-value targets — researching the recipient's role, their manager's name, recent company announcements — but this personalisation is costly and limits the attack to small numbers of high-priority targets. AI systems can perform this personalisation at scale: researching and incorporating personal details for thousands of targets in the time a human operator would research one.
Business Email Compromise — a category of fraud where attackers impersonate executives or vendors to trick employees into authorising fraudulent transactions — accounted for more than US$3 billion in losses reported to the FBI's Internet Crime Complaint Center (IC3) in 2025. AI can make BEC attacks both more convincing (better personalisation and language quality) and more accessible (lower operator skill requirement), although the IC3 notes that not all BEC tactics are AI-enabled; businesses reported more than US$30 million in losses to BEC scams involving AI in 2025.
Technical Deep-Dive
Modern AI phishing attacks operate across three layers. The targeting layer uses OSINT (Open Source Intelligence) tools — scanning LinkedIn, company websites, social media, and leaked data — to build detailed profiles of targets. The generation layer uses an LLM to create personalised email content that matches the target's expectations: referencing real projects, real colleagues, real deadlines. The delivery layer bypasses email authentication controls through compromised legitimate email infrastructure rather than obviously spoofed domains.
Email authentication standards (SPF, DKIM, DMARC) are effective against naive domain spoofing — they prevent emails from fake domains from appearing legitimate. They are ineffective against attacks that use compromised legitimate email accounts or that register look-alike domains that pass visual inspection.
Voice cloning for vishing (voice phishing) adds another dimension. CEO fraud — where employees receive phone calls apparently from their CEO requesting urgent wire transfers — has historically been limited by the requirement to impersonate the CEO's voice convincingly. Current voice cloning technology, trained on short audio samples available from public videos, can generate real-time voice synthesis that passes casual recognition by employees who know the target's voice.
The ASEAN Perspective
ASEAN businesses — particularly SMEs — are increasingly targeted by BEC attacks that exploit the region's cross-border business relationships. An Indonesian manufacturing company receiving what appears to be an email from its Singapore-based bank requesting account details, with personalisation that references recent transactions, is difficult for an employee to identify as fraudulent.
Verifying high-value transfer requests by telephone to a pre-registered number, rather than relying on email alone, is a widely recommended out-of-band control. This procedural control is more effective than technical controls against AI-generated phishing.
For ASEAN businesses of all sizes, the most effective countermeasure is not technical — it is procedural. Verifying requests for financial transactions through an independent channel (phone call to a pre-registered number, not the number in the email) eliminates the effectiveness of even the most convincing AI-generated phishing for the highest-risk transaction category.
RECATOOLS Verdict
The human-AI arms race in phishing will not be won by technical controls alone. AI generates attacks faster than detection rules can be updated, and personalisation defeats signature-based detection entirely. The durable defence is a combination of procedural verification (independent channel confirmation for financial transactions), awareness training that accounts for AI-generated content sophistication, and technical controls that reduce the attack surface.
For ASEAN businesses, implementing a simple policy — no wire transfers without independent telephone verification to a pre-registered number — removes much of the BEC attack risk, however convincing the phishing email is.
Frequently Asked Questions
It can be highly convincing: language models produce fluent, personalised messages without the errors that often gave older phishing away, so the content of an email alone is a weaker warning sign than it used to be.
Fraud where attackers impersonate executives or vendors in email to trick employees into authorising fraudulent wire transfers — with more than US$3 billion in losses reported to the FBI's IC3 in 2025.
Partially — SPF, DKIM, and DMARC prevent naive domain spoofing but are ineffective against attacks using compromised legitimate accounts or convincing look-alike domains.
Voice phishing — attackers impersonating executives or banks over the phone. AI voice cloning can now generate convincing voice synthesis from short audio samples.
Independent channel verification for financial transactions — phone confirmation to a pre-registered number before processing any wire transfer requested by email.
Correction, 18 October 2026: An earlier version described Business Email Compromise losses reported to the FBI as a global total, and cited click-through and testing results that we could not source; those statements and the headline have been revised. A typo in the list of email authentication standards has also been fixed.
View author profile → · Editorial policy
