AI fraud detection is now a brand problem, not just a bank's

What ASIC's 19,400 takedowns mean for ai fraud detection#

A fox does not need to see the whole field to know something is wrong with it. One dropped scent, one snapped twig in the undergrowth, and the whole hedgerow goes quiet. Australia's corporate regulator just had that moment, at scale, and it is worth every brand marketer's attention, not only every compliance officer's.

On 17 August 2026, the Australian Securities and Investments Commission published release 26-195MR, warning that scammers are using artificial intelligence to spin vast webs of deception. In the year to June 2026, ASIC removed 19,400 scam websites, social media ads and phishing hyperlinks, up 182 per cent on the 6,915 it took down the year before. Since the takedown programme launched three years ago, the running total has passed 33,400.

Break the number down and the shape of the problem sharpens. Phishing hyperlinks were removed at a rate of 5,476, up 279 per cent year on year. Fake investment platforms followed at 7,051 removals, up 151 per cent. Crypto scam sites climbed a comparatively modest 30 per cent to 3,106. Every category grew. None of them grew slowly, and that is the shape of a problem ai fraud detection budgets are scrambling to catch.

The detail that should worry a marketing department more than a fraud team is whose face the fakes wore. ASIC's release names Prime Minister Anthony Albanese, mining magnate Gina Rinehart, finance journalist Alan Kohler, economist Stephen Koukoulas, senator Jacqui Lambie and half a dozen other recognisable Australians whose faces and voices were stitched into deepfake videos promoting fake investment platforms. "AI is making investment scams more convincing and harder to detect," said ASIC Chair Sarah Court. "A simple online search is not enough to verify whether an opportunity is legitimate."

Losses tied to those campaigns reached $7.4 million by ASIC's own count, drawn from Scamwatch reports. That is a fraction of the wider picture. Australia's National Anti-Scam Centre logged $76.7 million in Scamwatch losses in the first quarter of 2026 alone, with investment scams accounting for $45.5 million of it, the single costliest category. Across the whole of 2025, investment scams cost Australians $837.7 million, up 7.8 per cent on the year before, inside total reported losses of $2.18 billion.

Why the logo, not just the ledger, is the target#

Read that release as a banking problem and you will miss the half of it built for you. Every deepfake ASIC describes needed a face people already trusted, a video format people already scroll past without thinking, and an advertising platform willing to run it. That is not a banking failure. That is a brand-marketing failure, wearing a compliance costume.

California's Attorney General made the platform layer explicit. In April 2026, Rob Bonta warned Californians that scammers were running deepfake investment scam ads across Facebook, Instagram and WhatsApp, cloning the faces of Cathie Wood, Kevin O'Leary and CNBC's Joe Kernen to lend the pitch borrowed credibility. "These scams have cost some people their life savings," Bonta said, "and typically use photos of well-known or trusted people."

Swap Wood, O'Leary and Kernen for your own chief executive, your own founder, the analyst your fintech quotes in every deck, and the exposure stops being abstract. A convincing forty-second clip of someone real, saying something they never said, beside your logo or product, is now cheap enough to produce and easy enough to place that the only defence left standing is detection, a fast takedown, and a public trail proving the platform, not the person, was compromised.

This is where the two secondary risks folkfox keeps tracking in cybersecurity marketing meet. AI deepfake scams borrow real people to sell fake platforms, and AI voice spoofing borrows real voices to sell fake urgency. Neither needs a security breach. Both need only an advertising account, a stolen clip, and a platform slow enough to let the ad run before anyone reports it. That gap between what a platform allows and what ai fraud detection actually catches is exactly where a deepfake investment scam ad buys its ninety seconds of unearned trust.

folkfox has already tracked this pattern once this month in Italy, where the Garante banned Mediaset's satirical deepfakes of journalist Enrico Mentana despite an on-screen disclosure label, a ruling covered in our piece on AI disclosure and deepfake labelling. A label is not a defence. Neither is hoping the platform notices first.

Meta's own advertising policies are under fresh scrutiny too, from a youth-safety trial folkfox covers separately, and the pattern rhymes: the platform sets the pace of harm, and the brand named in the fake ad pays the reputational bill.

What ai fraud detection now costs, and requires#

Here is the number that belongs in every 2027 budget conversation, not just every fraud team's. Deloitte's Center for Financial Services projects that generative-AI-enabled fraud losses in the United States alone will climb from around $12.3 billion in 2023 to as much as $40 billion by 2027, a 32 per cent compound annual growth rate. "Fake content has never been easier to create, or harder to catch," the firm warns. "As threats grow, banks can invest in AI and other technologies to help detect fraud and prevent losses."

Identity verification vendor Sumsub's Identity Fraud Report 2025-2026 makes the ai fraud detection budget case in finer detail. The global identity fraud rate eased slightly to 2.2 per cent in 2025, down from 2.6 per cent in 2024, but still above 2023's 2.0 per cent, because sophisticated, multi-layered fraud attempts surged 180 per cent over the same period. Deepfakes now sit among the top first-party fraud schemes Sumsub tracks, alongside synthetic identity and application fraud, and the firm expects agentic fraud, autonomous systems that script and carry out a full verification attempt end to end, to move from experimental probe to major wave through 2026.

None of that is theoretical for enterprise-facing brands. TechCrunch's reporting on deepfake fraud found 72 per cent of business leaders expect AI-generated fraud to be a top operational challenge in 2026, with 2025's deepfake attacks alone estimated to have cost organisations up to $1.5 billion. Deepfake fraud attempts roughly doubled in banking since 2023, and climbed sixfold in payments and sevenfold across gig-economy platforms over the same window. Incode founder Ricardo Amper put it plainly: deepfakes "have moved well beyond novelty" and "become a serious fraud weapon."

So what does buying ai fraud detection actually mean in practice, for a marketing or GRC team rather than a bank's core fraud desk? Three things, consistently: liveness checks confirming a face or voice is live and human rather than replayed or synthesised; provenance tools that watermark or fingerprint genuine executive and spokesperson media so an impostor clip can be flagged fast; and platform-level monitoring that catches an impersonating ad before it has run long enough to convert. None of the three is cheap, and none of them is optional once your executives are recognisable enough to be worth cloning.

The detection arms race nobody's winning yet#

Here is the part vendors selling ai fraud detection would rather brand marketers not read too closely. A peer-reviewed study in PMC tested detectors trained on the standard FaceForensics++ benchmark against DFDC and Celeb-DF instead, and found accuracy fell by roughly 22 to 28 percentage points, from the low nineties down into the sixties, evidence that many detectors are learning a dataset's own compression artefacts rather than anything generalisably fake about a fake.

Translate that into procurement language: a vendor's demo accuracy, measured on its own curated set, is not your accuracy, measured on whatever a scammer uploads to your ad platform next week. The UK's Financial Conduct Authority reached a similar conclusion from the regulator's side. Its Emerging Technology Horizon Scan 2026, published in June, names synthetic crime one of three headline themes reshaping financial services, warning that synthetic media "is becoming harder to tell apart from real content" and could "expose consumers and firms to new forms of fraud and deception."

The scale of the response is catching up, slowly. Detection vendors Veriff, Sumsub and Veridas told Biometric Update that deepfake detection checks are projected to climb from just over 3 billion in 2025 to nearly 10 billion by 2028, more than tripling in three years. Over six in ten fraud professionals in the UK (61.5 per cent) and the US (60.5 per cent) already report rising AI-driven attacks, a figure that reaches 69 per cent in Brazil.

None of this argues for giving up on detection. It argues for treating any vendor's accuracy number the way you would treat a stranger's investment tip: interesting, unverified, worth a second source before you act on it. The FTC's Voice Cloning Challenge is a useful signal of where the genuinely hard problems sit: its four winning entries covered synthetic-voice pattern detection, real-time deepfake and voice-clone flagging in two-second audio chunks, inaudible watermarking that scrambles an AI clone's training input, and voice authentication embedded as its own watermark. None of the four claims to solve AI voice spoofing outright. All four treat it as a moving target.

What a deepfake investment scam actually looks like in an ad account#

Strip the theatrics and a deepfake investment scam follows the same four beats every time: borrow a trusted face or voice, generate a short clip of them recommending a platform, buy inexpensive ad placement to push it past casual scrutiny, then route anyone who clicks toward a fake trading dashboard designed to accept a first deposit and stall every withdrawal after it. Recognising that pattern is the fastest way for a brand-side marketer to spot it happening to their own executive, weeks before legal or comms would otherwise hear about it.

Auditing your own exposure, before a regulator does#

A fox that waits for the hound to bark has already lost half its head start. The brands that come out of this cycle looking careful, rather than caught out, are running the audit now, not after the first fake clip of their chief executive turns up in someone else's ad account.

The story moved fast once ASIC published it. Within hours, financial-news watchers on X were relaying the regulator's core warning almost verbatim, which is itself a small lesson: a genuine regulatory release now travels at the same speed as the scams it describes, so a brand's own monitoring has to run on the same clock.

Five checks belong on that clock. First, catalogue every piece of public-facing video and audio of your named executives and spokespeople, the raw material any deepfake needs. Second, set a standing search across the major ad libraries for your brand and executive names, paired with words like investment, returns or trading. Third, ask any ai fraud detection or identity-verification vendor you are evaluating for accuracy figures from an independent, cross-dataset benchmark, not their own curated set, given what independent cross-dataset testing found.

Fourth, publish a plain, findable statement of what your brand will and will not ask customers to do, so a scam pitch that contradicts it is instantly suspicious rather than plausible. Fifth, brief your ad-platform and comms teams on a single reporting channel so a spotted fake reaches takedown within hours, not weeks.

This is squarely brand-strategy work as much as security work, which is why folkfox treats it that way. Our brand strategy practice builds the plain, quotable statements a fake pitch has to contradict, and our SEO and GEO team makes sure the genuine version of your executive's story is the one search engines and AI answers surface first, so a deepfake has less oxygen to spread in. Compliance teams already carry the KYC side of this argument, mapped in our earlier piece on deepfakes in gambling KYC risk assessment.

The regulators are not slowing down. ASIC's 182 per cent is a floor, not a ceiling, and every market watching Australia's numbers this week is quietly checking its own. The brands that treat ai fraud detection as a marketing discipline, not only a security line item, are the ones whose name survives the next wave of vast webs of deception with its trust intact. If you want that audit built rather than described, talk to folkfox.

Frequently asked questions#