Deepfake Threat Report, Week of 24 August 2026: Singapore Makes Platforms Verify Who Pays for Ads

Singapore orders Facebook, Instagram and TikTok to verify advertisers

On 18 August the Singapore Police Force announced three Codes of Practice under the Online Criminal Harms Act 2023, according to The Straits Times. The social media code, which the police text says takes effect from 17 August 2026, covers Facebook, Instagram and TikTok. Police described those services as posing the highest scam risk to users in Singapore; together they accounted for about 30% of scam cases in 2025.

The code requires platforms to verify advertisers by checking government-issued records, such as business registration records and identification documents, before any ad is published. It defines an advertiser broadly, to include agencies, influencers, affiliate marketers and anyone else who pays for or places an ad. Platforms must also refuse ads for financial services from firms that are not licensed by the Monetary Authority of Singapore, and remove live ads when there is reason to suspect a scam. Most obligations apply by 31 January 2027, and non-compliance is a criminal offence with fines of up to S$1 million.

The rule targets the step that deepfake investment scams depend on. The Australian Securities and Investments Commission said in its August warning that it has seen a sharp rise in scams using deepfake videos of celebrities and politicians, often starting from an ad on a well-known platform and leading to fake news articles and bogus trading sites. ASIC named Prime Minister Anthony Albanese as the most impersonated public figure in FY26 and linked $7.4 million in Scamwatch-reported losses to those impersonations. Our guide to investment persona scams covers how those funnels convert victims.

Advertiser identity checks would have stopped an anonymous buyer from placing those ads in the first place. They do not judge the creative itself, so a verified but compromised or complicit account can still run a deepfake endorsement. Platforms that want to meet the code in spirit will need to screen the faces and voices in ad creative as well as the identity of the person paying.

Meta ran deepfake ads for the Kromix nudify app

WIRED reported, and 9to5Mac summarised on 18 August, that Meta ran 32 ads for Kromix, an app marketed as an AI image styler that let users put real people's faces into pornographic video. The ads were targeted only at male users. One showed a woman closely resembling a prominent female US politician standing in front of a US flag, then cut to a pornographic clip with the same face.

The attack combined two weaknesses. The ads themselves carried non-consensual synthetic explicit media of an identifiable woman, which should have failed creative review. The app behind them also changed after approval: Apple said Kromix appeared to have added prohibited content and features after it passed initial App Store review. Apple removed the app, and Meta took down the campaign, after WIRED asked for comment. Apple also removed MaskAI, a face-swap app linked to problematic Meta ads in a separate WIRED investigation.

Face manipulation detection run on every ad video before it serves, and again on a sample of live ads, would have flagged a face swapped onto a pornographic body. Re-reviewing apps and advertisers after approval addresses the second weakness, since a one-time check at the door is what Kromix got past.

An AI-voiced bomb hoax closed a California high school

At about 2:07 a.m. on Tuesday 18 August, Yuba County Dispatch received what Wheatland police called a computer-generated, telephonic bomb threat against Wheatland Union High School. The caller claimed an explosive device on campus would detonate unless a ransom was paid. Superintendent Nichole Newman said the call used an AI-generated voice and fit a pattern of hoax threats against schools across the country, the Appeal-Democrat reported.

Nine explosive-detection K-9 teams from agencies including the Placer, Sacramento and Yuba County sheriff's offices, the California Highway Patrol and Beale Air Force Base searched the campus. They found nothing, and the school was declared safe at about 11:30 a.m. Police judged the threat not credible, but the district still delayed classes for its roughly 1,100 students.

A synthetic voice gives a hoaxer anonymity and lets one person place many calls without a recognisable voice. Detecting that a call is machine-generated would not change the duty to search a campus, but it adds evidence for investigators linking calls across districts. That is the same capability that protects contact centres and help desks from voice cloning attacks, where a synthetic voice is used to request a payment or a password reset rather than a ransom.

A congressman posted AI videos of his challenger

Spectrum News reported on 19 August that Rep. Derrick Van Orden of Wisconsin had been posting deepfake videos of his Democratic challenger, Rebecca Cooke, altered to make it appear she said things she did not say. Wisconsin Public Radio later identified one such post, from 4 August, showing an AI version of Cooke speaking into a microphone with an AI image of former President Joe Biden behind her.

The posts carry the words "Made with Grok Imagine," the xAI generator used to make them. Wisconsin law requires certain manipulated campaign content to carry the disclaimer "This video content generated by AI," which the videos did not include, though Spectrum News noted it was not immediately clear whether the videos meet the law's definition of regulated content. Cooke called the videos deception that should be illegal; a Van Orden spokesperson did not say why he posted them.

This case is a disclosure problem more than a detection problem: the author is known and the tool is named. Where detection matters is downstream, when clips are reposted without the label and reach voters, newsrooms or platform moderators who need to document what was altered and preserve evidence for a complaint or lawsuit.

What security teams should do this week

This week's deepfakes reached people through paid ads, an app store, a phone line and social posts. These are the controls worth reviewing now:

• If you run an ad platform, marketplace or creator programme that serves Singapore users, map the code's advertiser verification and MAS-licence checks against your onboarding flow before the 31 January 2027 deadline.
• Screen faces and voices in ad creative and user-uploaded promotional video for manipulation, and re-screen live ads and approved apps on a schedule rather than only at submission.
• Brief brand, communications and executive protection teams on deepfake endorsements: search for ads using your executives' faces, and agree a takedown route with each platform.
• Give call-handling teams, including emergency and facilities lines, a procedure for suspected synthetic voices: keep the recording, log call metadata and share it with investigators.
• For public-facing leaders and candidates, preserve original footage so a fabricated clip can be compared against what was actually said.

Where Neural Defend fits

For incidents like the Kromix ads and the Cooke videos, the work is investigative: establish what was altered, show the evidence and hand over a record that holds up. NeuroDetect is a forensic case workspace for that job. An analyst uploads an authorised image, video, audio file or document into a case; the analysis keeps model and version context, shows explainability heatmaps down to regions, frames and audio segments, and produces a downloadable forensic report with observations, uncertainty, provenance and analyst notes. The analyst makes the final call.

For checks that must run inside a product flow, such as verifying the person behind an advertiser account or screening the face in an ad video, NeuroVerify runs six signals in one API call, including Face Manipulation Detection, AI Generative Model Analysis and Visual Safety Screening, and returns a unified risk score with signal-level evidence in under 200 ms p95 on eligible image requests under defined test conditions. It does not analyse ID documents, so the government-record checks in Singapore's code need separate tooling. For voice threats like the Wheatland call, NeuroCall (Beta) detects deepfake voices on calls.

Neural Defend's detection models build on ARCAS 1B, its own deepfake detection foundation model. 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 measures still images only, and one of its authors also co-authored the ARCAS paper.