(Credit: © Gorodenkoff - stock.adobe.com)
A Reported Fake Security Warning Ad Stayed Up on Google
In A Nutshell
- Researchers found thousands of deceptive software ads in Google’s public ad archive.
- Scareware ads, which use made-up threats like fake virus alerts, were shown more than 16 million times.
- Ads making false promises, such as phone tracking or photo recovery, drew about 90 million impressions.
- Google judged most of the 20-plus ads reported as not violating its rules, and 41 ads tied to one malware-linked site stayed up.
An ad styled like a phone alert warns that “178 viruses were found on your Android phone,” in a translation from Polish. Few things send a thumb to the screen faster than a message like that. Researchers classified it as scareware, an ad that uses made-up threats to rush people into installing an app, the kind of ad Google’s own policies aim to catch. That one ad was shown on Google 1.1 million times, and scareware ads as a group were shown more than 16 million times across Europe.
Researchers at New York University and Radboud University in the Netherlands built an AI tool to hunt for these ads. It turned up thousands of deceptive software ads in Google’s Ads Transparency Center, a public archive that logs impressions for ads shown in Europe. Some promoted apps that app stores had already removed. Others led to websites that security services flagged as potentially malicious.
Google counts impressions, meaning each time an ad appears on a screen. The figures here show how often ads were displayed, not how many different people saw them or clicked. The sample covers software-related ads, not all of Google’s advertising.
AI Screening Caught Hundreds of Fake Security Warning Ads on Google
In a paper for the Proceedings of the 2026 ACM SIGSAC Conference on Computer and Communications Security, the researchers sampled 90,000 ads from that archive tied to software, mobile apps, and consumer electronics, running from early 2023 to March 2026. Counting every version of each ad, the collection reached roughly 188,000.
Hand review was impractical at that scale, so the team built a screening tool called AdLens. A small AI model read the text in each ad and ranked how closely it matched known scam patterns, such as fake account suspension notices. About 11,000 top-ranked ads moved on to larger AI models that examined image and text together, and a third model broke ties.
Three expert reviewers hand-labeled ads to check the system, and they agreed with one another closely. On the hardest calls, where the first two models disagreed, the tiebreaker scored about 92 percent on a standard measure that rewards catching deceptive ads without flagging honest ones.
Phone Number Tracking Ad Alone Drew 14.5 Million Impressions
Deceptive ads fell into three groups. Scareware, the smallest, plays on worries like viruses or a full phone, which the paper says are hard for users to verify. The team found 238 such ads from 54 advertisers. A French ad for a cleaning app claimed a browser was “70% damaged” after visits to adult websites.
Deceptive claims, the second group, are believable but false promises. These 3,346 ads from 279 advertisers drew about 89.9 million impressions, and a single advertiser ran nearly a quarter of them. Topping the list was an ad, active for more than two years, for an app promising to reveal anyone’s GPS location from a phone number. It collected 14.5 million impressions. Another told users it was their “last chance” to restore 284 photos deleted seven years earlier, drawing 2.8 million impressions.
Misleading design, the third group, hides who is behind an ad. A French ad showed only the word “Continuer,” meaning “Continue,” beside a green shield icon, and drew 1.6 million impressions. Reviewers on Trustpilot, a consumer review site, said the site behind it misled them into entering payment details for a paid subscription.
Google Left a Reported Fake Security Warning Ad Running
Of 147 apps promoted by flagged ads, 34, about 23 percent, had already been removed from app stores while their ads kept running. Among Google Play apps still available, file recovery tools and phone trackers were the most common, with installs reaching the hundreds of millions combined. Reviews of some of those apps complained of unauthorized charges, though the researchers call such reviews anecdotal.
Researchers checked 14,019 web addresses pulled from ads against four free security services that block dangerous sites. Eleven of them were flagged by at least two, though the researchers suspect one was a false alarm. Three of those were linked by security researchers, including Google’s own threat intelligence team, to malware or hacking operations, including a campaign that hid malware in a PDF editor.
Then came a test of Google’s “Report this ad” button. The researchers reported more than 20 ads, and Google judged most as not violating its rules. After one ad tied to the PDF editor campaign was reported, it came down, but 41 others pointing to the same website stayed up. For a scareware ad, Google said it could not review the ad at all, and the ad kept collecting impressions.
How many people installed a risky app or lost money remains unknown. What the study does document is a screening system with holes, where some ads survived even a direct report. Google reported blocking or removing roughly 1.7 billion ads in two enforcement categories in 2025. A system that removes that many and still lets a reported scareware ad stay up has a gap worth closing.
Disclaimer: This article is based on published research and is for informational purposes only. The study measures how often ads were displayed and cannot show how many people were harmed. As with all scientific research, findings may be subject to replication and further study.
Paper Notes
Limitations
Several boundaries apply. The ranking system reads text pulled from ad images, so it cannot catch deception that appears only in video or audio. The text-reading tool does not support non-Latin scripts such as Arabic and Cyrillic, though these made up fewer than 1.5 percent of ads. Because the data come from Google’s Ads Transparency Center, the sample includes only ads shown in at least one European Economic Area country, so deceptive patterns common elsewhere could be missed. Ads on Meta, TikTok, and other platforms are excluded, and the study targets three types of deception tied to software and mobile apps rather than all harmful advertising. The counts are described as lower bounds on how common each type is.
Impression figures measure exposure rather than confirmed harm, are estimated from midpoints of ranges Google provides, and cover only views registered in Europe. Domain flags came from four free security services and may include mistakes, which the researchers acknowledge for at least one domain. For misleading design, the test set held only three true examples, so its near-perfect scores should be read cautiously. Because the system relies on AI language models that follow instructions, it could be fooled by adversarial text crafted to evade detection. The reporting test used one EU-based account belonging to a senior coauthor.
Funding and Disclosures
Funding came from the National Science Foundation under Grant No. 2341206. Gunes Acar, Marwan Adnan Darwish, and Masoud Poorghaffar Aghdam were supported by the Netherlands Organisation for Scientific Research through a Vidi grant. The paper does not include a separate conflict-of-interest statement.
Publication Details
“AdLens: Efficient Detection of Deceptive Software Ads” was authored by Ritik Roongta and Rachel Greenstadt of New York University, and Marwan Adnan Darwish, Masoud Poorghaffar Aghdam, and Gunes Acar of Radboud University. It was published in the Proceedings of the 2026 ACM SIGSAC Conference on Computer and Communications Security (CCS ’26), held November 15 to 19, 2026, in The Hague, Netherlands. The paper runs 15 pages and is licensed under Creative Commons Attribution 4.0 International. Code and data are publicly released at github.com/Racro/ccs-2026-AdLens. DOI: https://doi.org/10.1145/3830454.3846781







