On November 2, 2026, WIRED published an investigation showing that a network of more than 400 connected apps and domains hid deepfake tools behind ordinary AI apps for pet photos, language lessons, fitness coaching, and similar services. Apple removed 19 apps after WIRED shared its findings with the company, while Google had already removed most of the apps before that. The key point is not any single app, but that a review mechanism was bypassed with a “pass review first, swap the content later” trick: what the app store saw was not necessarily what users ended up opening.
How they hid: pass review first, swap the content later
The investigation describes the method in three steps: the app is submitted for review with a harmless feature and appears to comply with store rules; after installation, it sends information such as language, referral channel, and device identifier to a remote server, which decides which version to show; and under certain conditions, the app loads a web interface offering tools to generate deepfake images and videos from photos users upload. Researchers also found a JavaScript file listing hundreds of the network's apps and domains, along with encryption keys and advertising trackers. WIRED and the research groups traced the domain trail to a Hong Kong-registered company called KeLing Technology Limited, which WIRED said did not respond to requests for comment; the attribution comes from the reporting and the researchers' tracing, and has not been confirmed by the company.
The scale figures also come from this investigation: in September 2026, researchers found 27 related apps on Google Play with more than 350,000 downloads combined; one of them, DiffAI, had more than 100,000 downloads and was removed by Google before WIRED contacted the company. An Apple spokesperson said the apps complied with the rules when reviewed, and that the developers changed the website content the apps load after approval. Meta said it had taken action against 11,600 advertising accounts and blocked advertising for 209 related apps. Apple's review guidelines, updated on October 16, already address non-consensual AI-generated intimate imagery and hidden features — but the tightening of the rules did not stop this network.
After the removals, the problem is not over
Removing 19 apps covered only the named portion. According to the investigation, the network added about 30 new app-and-website pairs per week at its peak, and similar ads were still appearing as the investigation wrapped up. For users, the more practical warning is this: approval in an app store does not mean everything an app loads has been reviewed, and apps of unclear origin, with vague feature descriptions, that ask for personal photos deserve particular caution. The platforms now face a concrete question: can server-side content swapping be detected while an app is running, rather than only after a press investigation? Marking AI-generated content so it can be recognized is another path already being tried, and the article on Claude text watermarking expanded to four models discusses how well such marking works in practice and how it can be bypassed.
