Meta's cloaking lawsuits target the oldest trick in front-end deception

The core technique is serving two versions of the same page

The most technically revealing part of Meta's February 26 announcement is its description of cloaking. As Meta puts it, a webpage linked to an ad "displays one version of its content to our ad review system, but shows different content to real users."

That is a front-end delivery problem before it is anything else. The scammer isn't hiding a payload deep in a binary; they are branching on who is requesting the page — likely on user agent, IP range, referrer, geography, or behavioral signals that distinguish an automated reviewer from a person tapping an ad. The clean version passes review. The dirty version reaches the target.

Meta says its "latest tools use AI to help us analyze cloaking and better detect ads that redirect to harmful websites," and that these tools help it reject ads faster and act more quickly on user reports. In practice that means Meta has to fetch and evaluate the page the way a real user would, not the way a crawler would — closing the very gap the cloaking relies on.

What the Vietnam case actually did to users

The lawsuit against Vietnam-based Lý Văn Lâm connects cloaking to a concrete consumer harm. Ads advertised "deeply discounted items from well-known brands, such as Longchamp," in exchange for completing a survey.

Users who engaged were redirected to sites asking for credit card information to buy items they never received — and their cards then "incurred unauthorized, recurring fees," which Meta names as subscription fraud. So cloaking got the ad approved; the survey-and-checkout flow harvested the card; the recurring charge was the actual business model.

For this fight to be efficient, we need to rely on active cooperation between all stakeholders, including intermediaries. We are happy that Meta takes action and demonstrates such cooperation. ~ Maison LongchampMontana Labs

Celeb-bait and the 500,000-face protection list

The China case against Shenzhen Yunzheng Technology Co., Ltd involves celeb-bait: misusing images of public figures to lure people, in this instance into fake "investment groups" targeting the US and Japan among other countries.

Meta says it now protects the images of "more than 500,000 celebrities and public figures around the world." That number implies a reference-matching system running against ad creative — detecting when a protected face appears in an ad and flagging it. But Meta is candid about the limit: "because scam ads are designed to look real, they're not always easy to detect." A legitimate celebrity endorsement and a fraudulent one can share identical surface signals, which is why the detection layer alone was never going to be sufficient.

The enforcement-evasion market is the harder target

Beyond the two lawsuits, Meta issued cease-and-desist letters to eight former Meta Business Partners selling "un-ban" services, account restoration, and "renting access to trusted accounts that helped clients evade our enforcement systems."

This is the part that should concern anyone building a review pipeline. It describes a secondary economy that exists specifically to defeat enforcement — reputation laundering through aged, trusted accounts. Meta says it is "reviewing our Business Partner ecosystem" and "enhancing our vetting methods," an admission that the trust it extends to partners was itself being monetized against it.

The implication: detection and litigation now have to move together

What makes this announcement specific is the pairing. Meta lists technical actions — suspending payment methods, disabling accounts, blocking domain names, and sharing indicators with industry partners — alongside actual lawsuits against named entities. It also cites the December 2025 Royal Thai Police operation that removed 59,000 accounts, Pages, and Groups and produced six arrest warrants.

The lesson for teams running ad or content review at scale is that cloaking cannot be solved purely at the model layer, because the adversary controls what the model sees. When the front end is deliberately adversarial, detection buys time and evidence; identifying the operators and pursuing them — through law enforcement or the courts — is what raises the cost of running the scheme again.