Google uses Gemini to stop maldumping and block millions of malicious ads

Google has recognized that it is increasingly using its Gemini models to identify and block malicious ads within its advertising networks, as scammers refine techniques to overcome traditional controls. In its latest public report the company details striking figures: during 2025 it withdrew or blocked about 8.3 billion ads and suspended almost 25 million advertisers' accounts, of which about 602 million were related to scams, as explained in its own blog on advertising security Keerat Sharma, responsible for privacy and security in Ads.

The maldumping - the use of advertising spaces to distribute malware, induce cryptomoneda theft or direct to phishing pages - has been a hydra head for platforms like Google for years. The attackers buy spaces and creative designs that imitate legitimate companies and services, and combine techniques of cloaking and readdresses to make the URL or appearance look confident, even showing domains related to Google or official download and authentication pages.

Reports from researchers and specialized media have documented specific campaigns that illustrate the variety and sophistication of the problem. Examples include ads that lead to forged pages to steal access to Google Ads accounts, ads used to distribute troyanized software by posing as popular utilities such as authentication managers or package installers, and ads that simulate being cryptomoneda platforms to empty visitors' money. Several press releases and press analyses have followed the lead on these campaigns, such as those covered by BleepingComputer and by security firms that show malicious creative that they get to impersonate Google's own page ( Malharebytes has published research in this regard).

What changes now is the scale and speed with which criminals can mount convincing operations: Google states that malicious actors are using generative artificial intelligence tools to produce mass misleading ads and create campaign assets much faster. The platform's response has been to incorporate Gemini into the control processes, not only to analyze the textual content of the pieces, but to cross billions of signals ranging from the advertiser's behavior and account history to campaign patterns and the apparent intention behind the announcement.

Before, detection systems were largely dependent on the analysis of key words and predefined rules; now, according to Google, models can evaluate much more complex contexts and automate the identification of suspicious campaigns before they are published, or block them in real time. This automation has also made it possible to speed up the processing of the reports of users who do manage to pass the initial filters, and to reduce the amount of accounts suspended by error: the company claims a decrease in the wrong suspensions of the order of 80% thanks to the greater accuracy of its models.

In the United States, for example, Google removed about 1.7 billion ads and suspended 3.3 million advertiser accounts in 2025, and points to violations such as the abuse of the advertising network and the supplanting of identity as the main causes of intervention, in accordance with its advertising policy rules ( Google Ads policy).

However, relying on IA to mitigate maldumping also poses challenges. The models are as good as the data and defense strategies that are fed to them; the attackers, on the other hand, quickly adapt their tactics - for example, by introducing layers of ofuscation, rotating domains, and using redirections and intermediate pages to make it difficult to track and block automated. In addition, the involvement of content generators implies that the volume of creative and forged pages can grow exponentially if it is not accompanied by equally agile controls.

For users and advertisers there are two practical readings. On the one hand, the good news is that the large platforms invest in advanced defence tools and that these systems are holding back on a large scale many abusive ads. On the other hand, the persistence of the problem recalls that no filter is infallible: it is necessary to keep classic precautions such as checking URLs and certificates, distrust of credentials requests on unexpected pages, use strong authentication and review permissions and downloads with security tools.

Google says it plans to expand Gemini's use to more advertising formats and compliance systems, with the intention to intercept malicious campaigns at the time of its shipment rather than react later. Meanwhile, the battle between defenders and attackers will continue to depend on both new technologies and human review processes and the common digital sense of users.

If you want to deepen the data and policies that Google publishes on this subject, you can consult its advertising security report on the official Google Ads blog Here. and for examples of specific campaigns and technical analysis there are research and journalistic articles in specialized sites such as BleepingComputer and in publications of cybersecurity firms as Malharebytes.