No malicious attachment. No exploit. No unusual process running on any endpoint. Nothing for your antivirus, your EDR, or your firewall to flag — because nothing in the attack is technically "wrong." Just a payment instruction that looks, in every observable way, completely legitimate. This is the category of fraud that's grown the fastest, and it's the one most security stacks are structurally unable to see, because they're built to catch bad code, not bad context.
This isn't a new problem. It's an old one, running at AI speed.
Business email compromise isn't new. Between October 2013 and December 2023, the FBI's Internet Crime Complaint Center recorded more than 305,000 BEC incidents worldwide, totaling over $55 billion in losses — including more than $20 billion and 158,000+ victims in the US alone. Losses were already climbing roughly 9% year over year before the current wave of AI tooling existed.
What's changed is speed and believability. AI-powered BEC specifically — attacks that use generative AI to craft the pretext — accounted for an estimated $2.77 billion in losses across more than 21,000 incidents in 2024 alone. Same underlying scam. Dramatically faster to produce, and dramatically harder for a trained employee to spot, because the tell-tale signs of a rushed, generic scam email are exactly what AI generation removes.
When the "proof" is a video call, and the video call is fake
In one of the most-cited cases of this shift, a finance employee at a multinational firm authorized 15 separate transfers totaling $25.6 million after joining a video conference with people who appeared to be senior colleagues, including the CFO. The employee had an initial instinct that something was off — and was talked out of it by a live call with synchronized facial movement and realistic voice replication. Every participant on that call except the employee was an AI-generated deepfake.
That's the endpoint of "context-based" fraud: not a suspicious link, but an entire fabricated context — a familiar face, a familiar voice, a normal-sounding meeting — built specifically to survive the one check most companies actually rely on: does this look and sound like the person it claims to be?
Why malware-focused defenses have nothing to catch here
Endpoint detection, antivirus, and email filtering are built around a specific assumption: that an attack involves running malicious code, exploiting a vulnerability, or delivering a payload that behaves abnormally once it lands. BEC and deepfake-enabled fraud do none of that. The "attack" is a human being convinced, through entirely legitimate-looking channels, to approve a transaction they were authorized to approve anyway. There's no signature to detect, because there's no malware — the entire attack lives in content and context, which is exactly the layer generative AI is now good at fabricating convincingly.
This is a verification gap, not a detection gap. And it needs a different kind of control.
The fix: verify out-of-band, independent of how convincing the request looks
The only check that reliably holds up against a perfectly convincing pretext is one that doesn't rely on judging how convincing the pretext is. Concretely: before any payment goes to a new or changed account, verify the payee's details — account number, routing information, the full set, not just the name — against a previously confirmed record, through a separate channel from the one the request arrived on. If the details match exactly, it proceeds without friction. If anything doesn't match — a new payee, a changed account number, anything — it gets a human out-of-band confirmation before money moves, regardless of how legitimate the request looked, who appeared to be asking, or how urgent it seemed.
The deliberate design choice here: this check has to be deterministic, not a judgment call — AI-generated fraud is specifically built to win judgment calls. A yes/no match against a verified record can't be talked into ignoring itself the way a person on a very convincing call can.
Where this stands today
CyberFence's Fraud & Payment Verification is currently in private pilot — we're building and testing it deliberately rather than shipping it before it's genuinely reliable, given what's actually at stake when it's wrong. It's not yet generally available.
Sources: FBI Internet Crime Complaint Center (IC3) BEC data, 2013–2023; industry reporting on AI-powered BEC losses, 2024; widely reported multinational deepfake video-conference fraud case.