✨ AI Summary
This article examines how artificial intelligence has transformed business email compromise (BEC) phishing from mass-volume attacks to precision operations, with particular focus on multi-persona campaigns where AI coordinates multiple fake identities simultaneously. It documents specific attack patterns observed by Proofpoint in 2026, explains why traditional phishing awareness training is now ineffective, and provides concrete technical and procedural defenses that remain reliable against AI-generated attacks.
Key insights
- AI shifted phishing from volume to precision gameTraditional phishing sent millions of generic emails with 0.1-1% click rates; AI-powered phishing targets 50 high-value individuals with 3-5x higher click rates (40% observed), according to IBM X-Force 2026 data. The economics fundamentally changed: fewer compromised credentials, but those credentials belong to people with admin access, financial authority, or privileged data access, making each successful compromise significantly more damaging.
- Multi-persona BEC campaigns build overwhelming social proofProofpoint's 2026 briefing documented campaigns where AI creates three coordinated fake personas—a CFO lookalike, legal contact, and supplier—that establish relationships over weeks before requesting payment. Each persona maintains consistency and remembers previous conversations, creating corroborating communications that defeat the single-source verification instinct that traditional BEC triggers. The victim receives reinforcement from three apparently independent sources, making the social proof overwhelming.
- Grammar errors and generic greetings are dead signalsAI phishing produces writing indistinguishable from legitimate corporate communication and personalizes each email with the target's real name, job title, company, and recent activity. These were traditional detection signals taught in phishing awareness training, but AI eliminates both. The article identifies this as the core reason why traditional training now teaches detection of signals that sophisticated attackers specifically no longer produce.
- DMARC protects against exact spoofing but not lookalikesDMARC enforcement (p=reject) blocks exact domain spoofing where attackers forge company.com as company.com, providing technical reliability against that specific vector. However, it does not block lookalike domains like c0mpany.com or company-corp.com, which AI campaigns routinely create. The article notes that DMARC combined with lookalike domain monitoring provides more complete coverage than either control alone.
- Out-of-band verification defeats all AI BEC attacksThe article identifies calling the real CFO on their known phone number—a number held independently and not obtained through the compromised email thread—as the single procedural control that stops multi-persona campaigns. AI cannot intercept a phone call to a number the target independently holds, making this procedurally reliable regardless of how convincing the email or the coordinated personas appear. The sophistication of the attack is irrelevant once this control is enforced.
- Fragment-based QR code attacks split malicious URLsProofpoint documented a 2026 technique where multiple QR codes in a single email each encode a fragment of a malicious URL, bypassing email security tools that scan hyperlinks but not QR code payloads. No single QR code contains the full malicious URL, so each fragment passes reputation checks independently. This represents a targeted attack on the detection layer rather than the user layer, specifically circumventing technical controls.
- Training must shift from detection to process adherenceThe article states that Gartner's 2026 cybersecurity trends report specifically noted GenAI is breaking traditional security awareness approaches. Updated training should remove lessons about spotting grammar errors and urgency signals, and instead teach finance teams to follow payment verification processes regardless of email quality: any new financial request requires out-of-band verification, any bank detail change requires phone confirmation, and new payment recipients require a second approver. The mental model must shift from 'I can spot a phishing email' to 'I follow the process regard
- AI enables OSINT at scale and voice cloning follow-upAI can research 10,000 targets in the time a human researches 10, using LinkedIn posts, company announcements, and press releases to identify high-value targets and create realistic pretexts. The article notes that voice cloning enables real-time phone follow-up using an AI-cloned voice of the supposed sender, and multi-channel coordination attacks across email, LinkedIn, WhatsApp, and phone simultaneously. These capabilities combine to create coherent, multi-week relationship-building that would be impossible at scale without automation.
Bottom line: The most effective defenses against AI phishing are procedural controls—out-of-band verification via phone call to a known number, dual approval for payments, and mandatory confirmation for bank detail changes—because these work regardless of email quality and cannot be defeated by AI. Technical controls like DMARC enforcement and phishing-resistant MFA provide necessary complement, but training must urgently shift from teaching people to detect suspicious content to teaching them to follow fina
