ADEX, the Limassol-based anti-fraud and traffic-quality business of AdTech Holding, said on October 27, 2026 that artificial intelligence played a significant role in 52% of the advertising fraud schemes it identified between January and September 2026. The figure comes from the company's State of Ad Fraud 2026: ADEX Trust & Safety Report, which draws on aggregated moderation and verification data from eight advertising platforms covering 1.2 million campaigns and 4.9 million creatives. Alongside the measured findings, ADEX makes a forecast: in 2027, fraudsters may increasingly go after the AI agents that browse, compare and buy on behalf of consumers.

In Short

A company that screens ads for scams looked at nine months of campaigns sent to eight ad platforms and says AI helped build about half of the fraud schemes it caught. That matters because AI makes fake news pages, fake celebrity endorsements and round-the-clock scam chats cheap to produce, so bad ads look more convincing and come back faster after being rejected. The company also predicts that next year scammers will try to trick the AI assistants people use to shop, which would mean ads get checked for what they tell a machine as well as what they show a person.

What the 52% figure counts

The headline number is not a share of campaigns, ads or advertiser accounts. According to the methodology note at the end of ADEX's release, "AI involvement is calculated across distinct fraud schemes, rather than individual campaigns or accounts." One scheme can span many campaigns and many creatives. The release does not say how many distinct schemes sit behind the 52%, and it does not explain how campaigns were grouped into schemes in the first place.

The wording also shifts between ADEX's own documents. The release says AI "played a significant role" in 52% of schemes. The report's one-page summary describes the measure differently: "About half of the abuse schemes ADEX observed in January to September 2026 showed indicators of AI-generated or AI-assisted assets." Neither text says what counts as an indicator, where the line for "significant" sits, or how often the detection might misfire. Those are not trivial gaps. Telling an AI-generated landing page from a template-built one is a judgment call, and the report gives no error rate for it.

Then there is the sample. The release refers to "eight advertising platforms using ADEX," while the report cover reads "8 ADEX clients." None of the eight is named. ADEX belongs to AdTech Holding, which also operates the PropellerAds advertising platform, and neither document says whether PropellerAds is among them. What the data describes, then, is the population of advertisers and fraud operators that tried to run campaigns on those eight platforms over nine months. It is not a measure of fraud across the advertising market as a whole.

According to the report, the data was "collected and analyzed by the ADEX analytics team based on moderation and verification data from ad platforms that use ADEX, and compared with 2025, where comparable data is available." Which metrics had a 2025 baseline is not specified. The release does not mention any independent review of the data, and ADEX sells the moderation and detection products whose output makes up the dataset.

Inside the AI-involved schemes

Among schemes where AI was involved, according to ADEX, bots impersonating humans appeared in 64%, cloned brands and fake endorsements in 53%, and deepfakes in 11%. The categories overlap, which is why the three shares add up to 128%. "The categories overlap, as fraudulent schemes often tend to combine several AI-powered techniques," the release states.

These shares apply to the AI-involved subset, not to all fraud. By PPC Land's calculation, applying them to the 52% headline figure means bots impersonating humans featured in roughly a third of all schemes ADEX identified, cloned brands and fake endorsements in a little over a quarter, and deepfakes in under 6%. Both inputs are rounded, so the derived figures are approximate.

What does "bots impersonating humans" mean here? The release does not define it. Its own description of a modern investment scam features a chatbot that "can then handle conversations around the clock, responding fluently in the victim's language," which points to automated conversation agents posing as people. Whether the category also covers automated traffic posing as human visitors - the kind ad buyers know as invalid traffic - is not stated.

Brand cloning, according to the report, appears across investment and blockchain scams as well as fraudulent e-commerce stores. Attackers also imitate familiar technology brands to distribute fake security alerts. ADEX names no brands and gives no breakdown of schemes by vertical.

The same scheme, a better finish

ADEX frames the change as one of production cost rather than method. A fake investment campaign of a couple of years ago, according to the release, "usually meant a hastily assembled landing page, maybe a celebrity photograph paired with a fabricated quote," with a small team working potential victims through private messages. Now, the company says, AI can produce a polished imitation of a legitimate news outlet or financial institution "that would look almost flawless to the naked eye," and a chatbot can take over the conversation.

Sergey Martianov, Chief Product Officer at ADEX, put it this way: "The scheme itself has not changed in years: borrow a trusted face, promise big returns, move the victim into a private chat. What AI changed is the finish. The page looks real, the video looks real, and the person in the chat never sleeps and never makes a typo."

The economic argument follows. "A set of agents now stands in the place of a team of content makers, designers and developers," according to the release, which adds that fraud content "is super cheap to produce and even cheaper to scale." No cost figures accompany the claim.

Rejected campaigns that come back

One operational detail stands out for anyone who runs an ad review queue. ADEX researchers, according to the release, observed cases in which revised versions of rejected campaigns returned within hours, sometimes on the same day. The company does not report a standard turnaround time, saying it varies considerably between schemes. No count of such cases is given, nor any distribution of how long resubmissions took, so the finding is best read as an observation rather than a measured rate.

Six shifts in the moderation data

The 52% figure is one of six changes listed on the report's summary page under the heading "2026 in one page: six things that changed this year." Read together, they describe a moderation workload that is shrinking in volume and growing in difficulty. According to the report, fewer policy-violating campaigns were submitted for moderation in 2026, "but the ones that were submitted were more complex." The operators behind them "moved away from high-volume submissions toward fewer, better-disguised campaigns, and generative AI played a growing role in how those campaigns were built."

  • Cloaking - showing reviewers one page and real visitors another - remains the top reason platforms suspend advertisers. Its share, however, fell from about three quarters of suspensions across 2025 to under half in the third quarter of 2026. The report's headline for this item is "Cloaking is losing its monopoly."
  • Malicious iOS scripts appeared as a separate reporting category and reason for platform suspensions in the third quarter, accounting for about one in five suspensions that quarter.
  • Copyright violations grew from 8% of campaigns rejected at platform moderation in the first quarter to more than a fifth by the third.
  • Adult content fell from 42% of rejections in the first quarter to 14% in the third, the largest drop among rejection reasons.
  • Messenger-based phishing, observed on services including WhatsApp and Telegram, became a separate reason for platform suspensions, with the persuasion taking place inside the chat. The report adds that "the conduct described is third-party abuse of those services; nothing in this report is a statement about the services or their operators."

Two cautions apply. The summary gives shares, not absolute numbers, so a falling share can reflect fewer cases of one type or more cases of another; cloaking's decline, for instance, coincides with a new iOS category taking about a fifth of suspensions. And the denominators differ. Cloaking and iOS scripts are measured against suspensions, while copyright and adult content are measured against rejected campaigns. The two are not interchangeable.

The cloaking trend is relevant beyond ADEX's eight platforms. The technique has been central to recent malvertising cases. When Confiant and independent researcher Michael Steele documented the SourTrade campaign in July 2026, landing pages fingerprinted each visitor, sending real victims to a malicious "money page" while researchers and bots saw a harmless "white page". That campaign ran across 12 countries, with evidence pointing to Google Ads, Meta and X advertising infrastructure.

The 2027 forecast: when the visitor is an agent

ADEX's second argument looks ahead, and the company is careful to label it. According to the release, the threat to AI shopping agents is "a forecast rather than an attack category currently measured in its moderation data." The 52% figure and the observed techniques come from 2026 data; the agent scenario does not.

The mechanism ADEX highlights is prompt injection, in which "malicious instructions are embedded in content an AI agent consumes while performing a legitimate task." According to the release, such instructions "could appear in invisible page elements, metadata, product images, or text addressed directly to AI assistants." ADEX also expects attackers to adapt cloaking for the new audience, serving different content to human visitors and AI agents, much as they already do to evade advertising moderation.

"In 2027, the most valuable user on the web may not be a person at all. If an agent is doing the shopping, someone will try to scam the agent. The industry needs to start checking ads for what they say to machines, too," Martianov said.

The release notes that prompt-injection techniques are already documented in security research, and that ADEX has previously examined how agents can be manipulated. That earlier work is a blog post published on September 24, 2026, titled "Do Not Mention This to the User: How an AI Agent Can Turn Into the Attacker's Delivery Boy," which compiled published cases of agents being steered toward malicious tools and packages.

The underlying technique has a public record. In August 2025, Brave's security team disclosed that Perplexity's Comet browser could be hijacked by hidden instructions placed in web pages through white text on white backgrounds, HTML comments or nearly invisible text in images. The same month, Anthropic said its tests of Claude for Chrome across 123 cases and 29 attack scenarios produced a 23.6% attack success rate without mitigations, falling to 11.2% with them. What ADEX adds is the advertising angle: the expectation that ad creatives and landing pages, not only ordinary web pages, will become a delivery route.

Shopping agents are no longer hypothetical either. On September 20, 2026, Amazon said it had cut Meta's Muse agent off from buying on Amazon.com, citing among other complaints that the agent does not identify itself when browsing.

Good agents, bad bots

The expected growth of shopping agents creates a classification problem that ADEX spends a section of its release on. An AI assistant that compares prices or buys something for a real customer "is doing a legitimate job, even if no human ever touches the website," according to the company. Blocking every automated visitor would turn away paying customers; admitting every one would let fraud through.

ADEX's answer is to drop the "bot or human?" question in favour of three others. Can the visitor prove who it is? Does it behave the way its stated task says it should? And do its sessions end in something real, such as a purchase or a booking? The company adds that platforms should watch for websites that show AI agents one page and people another.

Each of those questions maps onto work already under way elsewhere. Identity is the territory of Web Bot Auth, the protocol that lets automated clients sign their requests cryptographically. Adoption remains thin: Arkose Labs reported on October 20, 2026 that only 2.3% of the AI agent sessions it classified as legitimate carried a verified signature. Behaviour matching the stated task is close to what DataDome calls "continuous intent verification," which it brought to Experian's Agent Trust programme on October 1, 2026. Experian had launched that programme in April with Visa, Cloudflare and Skyfire, linking verified consumers to the agents that act for them.

The release also lists three checks for advertisers: watching for their brand in other people's ads, checking where their clicks actually end up, and asking technology partners how they tell a good agent from bad automation. It closes on a question of its own: "who, or what, is making the decision, and what told it to?"

Who ADEX is

ADEX - short for Ad Expert, according to its website - describes itself as part of AdTech Holding, "a group building advertising technologies since 2011." The website's footer lists Samoukale Enterprises Limited, with an address at Zinonos Rossidi 11 in Limassol, Cyprus. The release describes ADEX as "a global anti-fraud and traffic-quality platform" that "analyzes billions of impressions, clicks, and conversions."

The platform is sold as four products. ProQuality scores impressions, clicks and conversions in real time, with pre-bid and post-bid integrations. ProModeration handles the campaign review the report's data comes from, listing AI analysis of ads, landing pages and redirects, "cloaking and conditional content detection," custom policy rules and human review for complex cases. ProShield targets account, affiliate and lead abuse, while ProIntel offers threat feeds, infrastructure and actor profiling. The site claims sub-50ms processing and lists adtech, lead generation, e-commerce, SaaS and platforms, and social and streaming among the sectors it serves.

The homepage advertises 50 billion-plus monthly checks, more than $250 million in ad fraud flagged and more than 200 detection threat signals. A footnote describes these as "figurative and based on internal platform data (rolling 12 months)."

Why this matters for marketers

For brands, the most direct exposure is cloning. If more than half of AI-involved schemes in ADEX's data borrow a brand or a face, then legitimate advertisers are the raw material for a large share of the fraud, whether or not they ever buy on the eight platforms in the sample. The pattern is familiar. Google tightened its Misrepresentation policy in early 2024 to target ads impersonating public figures, brands or organisations, with enforcement starting in March 2024. More recently, on October 7, 2026, analytics engineer David Vallejo described being phished through a Google search ad that led to a copy of the Google Cloud Console homepage - the kind of technology-brand imitation ADEX describes.

Scale on the largest platforms gives some sense of proportion. Google's 2025 Ads Safety Report, published in April 2026, said the company blocked or removed 8.3 billion ads and suspended 24.9 million advertiser accounts in 2025, including 602 million ads and 4 million accounts for scam-related violations. Against those numbers, 1.2 million campaigns across eight platforms is a modest dataset - though ADEX's unit of analysis, the scheme, is different from Google's counts of ads and accounts.

Vendor fraud figures rarely line up, and this report will not change that. DoubleVerify said in July 2026 that fraud and invalid traffic fell 41% in North America and 45% in EMEA year over year, while also finding that AI bots generated 15% of clicks across unprotected media in the campaigns it studied. DoubleVerify measures impressions and clicks on its clients' media. ADEX counts schemes in moderation queues. A falling fraud rate in one and a rising AI share in the other are not contradictory, because they measure different things.

The agent forecast matters more to performance marketers. Click-based metrics, conversion data and bid algorithms all assume that a visitor is either a person or a bot to be filtered out. A legitimate agent buying on a customer's behalf breaks that assumption, and an agent manipulated by a landing page breaks it differently. ADEX offers no data on how often that happens at present - by its own account, it does not yet measure it. Whether its 2027 prediction holds will depend in part on how quickly identity standards spread among the agents themselves.

Timeline

Summary

Who: ADEX, the anti-fraud and traffic-quality business of AdTech Holding, based in Limassol, Cyprus, with comments from Chief Product Officer Sergey Martianov. The data comes from eight unnamed advertising platforms that use ADEX.

What: State of Ad Fraud 2026: ADEX Trust & Safety Report, which finds that AI played a significant role in 52% of the ad fraud schemes ADEX identified. Among those AI-involved schemes, bots impersonating humans appeared in 64%, cloned brands and fake endorsements in 53% and deepfakes in 11%, with overlap. The report also tracks cloaking falling below half of suspensions, new iOS script and messenger phishing categories, and rising copyright rejections. It forecasts, without measured data, that fraudsters will target AI shopping agents in 2027. The release does not disclose the number of schemes, the platforms' names or error rates for its AI detection.

When: The findings were released on October 27, 2026. The data covers January to September 2026, with comparisons to 2025 where ADEX says comparable data exists.

Where: Moderation and verification queues at eight ad platforms using ADEX, covering 1.2 million campaigns and 4.9 million creatives. The release names no countries or markets.

Why: ADEX argues that AI has made established scams cheaper to produce, more convincing and faster to resubmit after rejection, and that the next shift will be aimed at the AI agents consumers use to shop. For brands, the findings point to cloning of their names and faces; for platforms and performance marketers, to a coming need to separate legitimate automated customers from manipulated or malicious ones.