Deepfake Threat Report, Week of 14 September 2026: First TAKE IT DOWN Act Sentence Lands

First TAKE IT DOWN Act sentence: 15 years for AI-generated forgeries in Ohio

On Tuesday 8 September 2026 a federal judge in the Southern District of Ohio sentenced James Strahler II of Columbus to 15 years in prison. He had pleaded guilty in April to cyberstalking, producing obscene visual representations of child sexual abuse, and publishing digital forgeries. U.S. Attorney Dominick S. Gerace II said his office believes Strahler is the first person in the country convicted under the TAKE IT DOWN Act, the 2025 federal law that criminalises publishing non-consensual intimate images, including AI-generated ones.

According to Spectrum News, investigators found more than 24 AI platforms and over 100 web-based AI models on his phone. Between December 2024 and June 2025 he targeted at least six women with real and AI-generated nude images, messaged their mothers demanding nude photographs, and threatened to release images of their daughters. He also used the faces of minor boys from his community in AI-generated sexual content and posted more than 700 images to a child sexual abuse website.

No single technical control would have stopped someone generating these images on a personal phone. What shortens the harm is fast reporting and removal: victims or their families flag the content, the hosting platform matches and removes copies, and investigators document which images are synthetic and which are real. That last step matters in court, because the charges for authentic and forged images differ.

Tech Transparency Project finds 332 Meta ads with AI-generated child abuse imagery

On 8 September the Tech Transparency Project (TTP) reported that 332 paid ads containing AI-generated child sexual abuse material passed Meta's review and ran on Facebook and Instagram during 2026. TTP counted 262 ads served in the U.S., 120 in Australia and 84 in India. Ad Library data showed reach of more than 29,000 users in the EU and 6,800 in the UK.

Most ads paired a clip of adult pornography with a photo of a child that AI had altered to depict a sex act. TTP identified photos of real children taken from social media, including a minor from a European royal family. Of the 332 ads, 298 pointed to image or video generator apps in the Apple and Google stores, and TTP verified that 182 of the linked App Store apps could nudify people. The ads ran from apparently fake pages, and several were placed through Meta's authorised resellers in China.

Meta said many of the ads predated detection systems it had recently launched. TTP reported that of 129 active ads it flagged, Meta rejected 73 as not violating its standards. The control gap is ad creative review: an image classifier tuned for nudity can miss an AI-altered photo of a real child, so creative screening needs manipulation and generative-model analysis, plus a check on what the destination app does after approval.

Sri Lanka-linked page network deepfakes Australian politicians for engagement

ABC News reported on 9 September that the research group Reset Tech had tracked more than 200 Facebook pages impersonating Australian political figures over eight months. Reset Tech linked them to an influencer business in Sri Lanka and believes the wider network runs 1,864 coordinated pages with a combined 5.9 million followers, also targeting France, Italy, the UK, the US, Canada and Mexico.

The pages mostly posted short Reels in which politicians such as Anthony Albanese, Pauline Hanson, Fatima Payman and David Pocock appeared to say inflammatory things on migration and religion. One fabricated clip showed Senator Payman appealing against deportation to Afghanistan, and another had Senator Pocock announcing an alliance with One Nation. Reset Tech found that only one in three posts carried an AI label and one in 33 was deleted. About 10 per cent of the pages were monetised through Facebook's creator payment scheme, which Reset Tech's Rys Farthing described as commercially driven foreign influence.

Meta said it was reviewing the pages and would act on any that broke its policies. The controls that would have caught this sit with the platform and with the people being impersonated: labelling of synthetic video at upload, detection of coordinated page networks, and a monitoring process that lets a political office send suspect clips for forensic analysis and request takedowns with evidence attached.

Apple adds sensor-signed Reference Image capture to the iPhone 18 Pro

Apple's 9 September announcement of the iPhone 18 Pro included Apple Reference Image, an opt-in mode in which the new main camera sensor cryptographically signs the data it captures. Private Cloud Compute turns that signed sensor data into a reference image that cannot be changed, and the Photos app can show it next to an edited version so viewers can see what was altered. Apple said SynthID support for identifying AI-generated or edited images will follow in a later software update.

This is a defensive development, and it has limits that security teams should plan around. Reference capture is not available in China at launch, and in the EU it is not available on the iPhone 18 Pro at launch, although users there can still develop and view reference images. It covers only photos where the owner switched the mode on. A signed reference image shows where and how a photo was captured; it says nothing about the millions of unsigned images that will keep arriving in claims, onboarding and evidence workflows.

Deepfake ad uses Alphonso Davies's father to promote an unlicensed casino

On Friday 11 September an Instagram account called Prime Spin Zone posted an ad with an AI-generated video of the father of Canada and Bayern Munich defender Alphonso Davies being arrested by Edmonton police, then released when officers learn his money came from gambling winnings. The Edmonton Journal reported on 14 September that the ad also used an AI likeness of the musician Drake and sent viewers to Oxibet, a Comoros-based site that is not on Alberta's list of 31 registered operators.

Davies's agent, Nedal Huoseh, called the ad lies and said the family does not gamble and wants to pursue those responsible. Alberta minister Dale Nally urged residents to use only approved operators. Ad review that screens faces in creatives for manipulation and checks landing domains against licensing registers would have flagged it before it ran.

What security teams should do this week

Four of this week's five items involve synthetic faces published at scale through social platforms and ad systems, and the fifth is a provenance control that will cover only a fraction of incoming media. These steps apply to trust and safety, brand protection and fraud teams.

• Write down a notice-and-removal procedure for synthetic intimate imagery, including who triages reports, how copies are matched, and how evidence is preserved for law enforcement.
• Screen ad and creator uploads for face manipulation and generative-model artifacts, and re-check destination apps and domains after approval to catch bait-and-switch changes.
• Monitor for impersonation pages of your executives and public-facing staff, and keep a forensic report template ready so takedown requests carry evidence.
• Accept signed provenance such as Apple Reference Image or C2PA Content Credentials where present, and keep detection running for the unsigned majority of files.
• Treat every detection score as one input to a policy decision, with human review for high-impact removals and legal referrals.

Where Neural Defend fits

NeuroDetect is the closest fit for this week's cases. Analysts upload an authorised image, video or audio file into a case, see explainability heatmaps down to regions and frames, and export a forensic report that records observations, uncertainty, provenance and model context. That report is the evidence a takedown request or referral needs.

NeuroVerify fits upstream, where faces enter a platform. One API call checks a face image, video or audio for manipulation and returns a risk score with six signal-level results, including Face Manipulation Detection, AI Generative Model Analysis and AI Content Compliance, in under 200 ms p95 on eligible image requests. Ad, creator and seller onboarding teams can run it on profile selfies and creatives before approval. On the public VendorBench-100 benchmark of 100 adversarial still images, Neural Defend ranked first of 36 systems with an MCC of 0.876 against 0.646 for the next vendor; that benchmark covers still images only.

No voice incident made this week's list; for phone-based impersonation, NeuroCall (Beta) detects voice deepfakes on calls.