Top Deepfakes of the Week: Aug 17–24, 2026

1. A "nudify" app bought ads on Meta and targeted a sitting politician

An app promoted as generating nonconsensual sexual images from a single uploaded photo ran dozens of ads on Meta's platforms this week. One ad opened with an AI image resembling a well-known female politician, then cut to explicit content built from the same face. The ad account was days old and had no followers — the kind of profile automated review systems are supposed to flag.

Apple pulled the app, plus a second one with the same function, from the App Store only after a reporter asked about it. Both companies have removed similar apps before. The pattern keeps repeating: an app passes review with clean functionality, then switches on the exploitative feature once it's live. Detecting this kind of manipulated AI-generated image after the fact is far harder than catching it before it reaches an ad queue.

2. A cloned voice over real footage cost Australians millions

Australia's securities regulator has now tied roughly AUD $7.4 million in investment scam losses to deepfakes of the country's most recognizable public figures, including the Prime Minister. The technique isn't a face swap. It's audio grafting: the video is completely real, and only the voice track is synthetic, cloned from a few seconds of reference audio and layered over real footage to promise guaranteed monthly returns.

That inverts the advice most people are given for spotting a deepfake. Watch for blinking, watch for lip-sync — none of that helps when the face on screen is genuinely the person's own. Only the audio itself is fake, which is exactly why voice cloning has become the fraud method of choice for scammers impersonating public figures.

3. A sitting congressman posted deepfakes of his own opponent

In Wisconsin, Rep. Derrick Van Orden has been posting AI-generated video that makes his Democratic opponent appear to say things she never said. State law requires an explicit on-screen disclaimer on AI-generated political content. The posts don't carry one — only a caption crediting the generation tool used.

This isn't an anonymous troll account. It's the official channel of a sitting member of Congress, in one of the most closely watched House races this cycle. Federal bills to regulate AI in elections have stalled repeatedly, so protection for voters currently depends entirely on which state they happen to live in — and whether that state's disclosure rules have any teeth.

4. AI-generated bomb threats closed schools in two states

Multiple districts in Georgia canceled classes on August 14 after receiving threats investigators described as computer-generated. A California high school closed on August 18 after an AI-generated voice delivered a threat with a ransom demand attached. Every building got searched. Every case, real or not, triggered the same emergency response.

That's the point. AI voice generation lets one person deliver a threat to multiple institutions at once, hard to trace and cheap to repeat. The ransom demand in the California case suggests someone is now testing whether AI voice can be used for extortion, not just disruption — a use case that's likely to keep showing up.

5. North Korean operatives face-swapped their way into US companies

A yearlong Wall Street Journal investigation, drawn from leaked data off the operatives' own machines, traced a North Korean IT worker cell that applied to more than 1,000 US companies in three months. AI wrote their resumes, assisted their interview answers, and swapped their faces in real time during live video calls — using stolen American identities. The FBI has since identified thousands of similar applicants; Treasury estimates the operation earns the regime hundreds of millions of dollars a year.

What makes this different from older identity fraud is the real-time part. A stolen name and a fabricated resume aren't new. Passing a live interview as someone else used to require an accomplice. Now it just requires software running during the call — which is exactly the gap that real-time deepfake detection for meetings is built to close, since a resume and a video call are no longer proof that someone is who they claim to be.

Honorable mentions

The pattern underneath all five stories

Every layer of moderation this week was built for a single point-in-time check — an ad review, an app store approval, a takedown after the fact — and every one of them missed the threat until someone outside the system flagged it. That gap is where scams like these live, and it's the same gap an AI scam and deepfake detector is meant to close for individuals and companies who can't wait for a platform to catch up.

The other shift is who's doing this. A member of Congress posting deepfakes from his own account. A state-sponsored cell sitting in job interviews. Coordinated threats hitting schools in two states at once. Synthetic media isn't confined to anonymous accounts anymore — it's being used openly, by organized and sometimes official actors, faster than regulation can catch up.

No single checkpoint catches all five attack types above — a nudify ad, a cloned voice, a mislabeled campaign video, a spoofed call, a face-swapped interview. That's why UncovAI builds detection as a full product suite rather than a single tool, so image, audio, video, and identity checks all live in one place instead of five different ones.

Frequently asked questions