Deepfake Threat Report, Week of 12 October 2026: Paris Hospitals File Complaint Over Fake Doctor Videos

The week's deepfake incidents at a glance

Four confirmed items shaped the week. Two involve impersonation for money: a hospital professor's likeness used to sell a supplement, and a parent's voice-clone kidnapping call that reached a national audience this week. Two involve political persuasion, where the legal question is labelling rather than fraud. Each item below is confirmed at the original institution or a news outlet that reported it directly.

AP-HP files a complaint over deepfake videos of a hospital parasitologist

AP-HP, the Paris public hospital group, and Sorbonne Université published a warning on 5 October about three Facebook videos that used the image and voice of Pr Renaud Piarroux, a parasitology specialist at Pitié-Salpêtrière hospital, without his consent. The videos promoted a product sold as ridding the body of parasites. AP-HP said no scientific data supports its efficacy or safety, filed a complaint and referred the videos to Meta.

Meta removed the three videos. AP-HP then reported that similar videos had appeared on Facebook again. Next reported on 6 October that one version runs about five minutes and imitates a TF1 news report, borrowing the identity of presenter Anne-Claire Coudray, and that copies circulated on pages and groups linked to alternative medicine and anti-vaccine views. Next linked the case to AFP's spring 2024 reporting on deepfakes of French public figures selling supposed diabetes cures.

The attack combines two borrowed sources of trust: a named hospital specialist and a familiar news format. Its audience, which Next described as mainly older people with pain or sleep problems, is less likely to question a doctor who appears to speak on camera. Repeat uploads mean every takedown is temporary unless the rights holder keeps watching.

Likeness monitoring across social platforms, paired with forensic analysis of each copy, is the control that would have caught it. A hospital or university that can show which regions and audio segments were synthesised, with model context, gives the platform and investigators a stronger complaint than a screenshot.

A cloned-voice virtual kidnapping case reaches a national audience

Fox News published an interview on 7 October with Deborah Del Mastro, a Navy veteran from Martinez, California, describing a fake kidnapping call. The incident itself happened in May 2026; ABC7 reported it on 26 May, and Del Mastro later testified before the US Senate Aging Committee. It is included here because the detailed account became public this week.

The caller rang from a local number, asked whether her daughter Sarah was prone to panic attacks, then played a voice that sounded exactly like Sarah crying. He claimed a Mexican drug cartel held her and demanded $20,000. Over about five and a half hours he kept the line open, moved Del Mastro and her husband between locations, gave them a cover story and had a second man demand more. They sent $5,400 in four transfers through MoneyGram and Western Union to Mexico before Del Mastro reached Sarah, who was safe at work.

The scheme is a classic family emergency voice scam with a cloned voice added to remove doubt. The decisive control is out-of-band verification: calling the person on a known number, or contacting someone who knows where they are, before any money moves. Wire services and banks are the second checkpoint, because a parent sending several urgent transfers to Mexico under instruction fits a known coercion pattern.

AI deepfake ads spread through the US midterm campaign

Al Jazeera reported on 9 October that candidates and their backers have spent more than $80 million on more than 164 AI-based ads this election season, according to the Wesleyan Media Project, with Republican candidates and backers behind 80 percent of them.

The clearest example is in the New York governor's race. Ads from Republican candidate Bruce Blakeman show Governor Kathy Hochul and New York City Mayor Zohran Mamdani in AI-generated scenes, including one of the two gardening together, though neither appeared. The ads show an AI disclaimer for a few seconds at the start. Blakeman called them satire and parody; New York Democrats filed a complaint with the state board of elections arguing they breach disclosure rules and mislead voters.

Al Jazeera noted that 31 states have passed laws requiring AI-made political ads to be labelled and that no federal law regulates them. For campaigns, election offices and newsrooms the risk is the same: a realistic clip of a real official spreads faster than a correction. The control here is rapid provenance and manipulation analysis on the clip in question, so a communications team can state what was generated and publish that finding while the ad is still circulating.

A Vermont PAC's AI video of Sen. Ruth Hardy tests the state's disclaimer law

VTDigger reported on 8 October that Flip Vermont Red, a PAC founded in September by Curtis Hier, posted an AI-generated video showing a realistic version of state senator Ruth Hardy placing a state regulation sign on a farm, forcing the farmer out and handing over a property tax bill. The only label is small text reading "AI made me do it." Hardy has been named or depicted in at least 11 of the PAC's posts.

Vermont's Act 75 requires a disclaimer on realistic AI-generated political content but exempts satire. The PAC's posts carry no disclaimers, and Hier said its synthetic media are neither deceptive nor fraudulent. The exemption has already decided one case: Attorney General Charity Clark investigated a separate undisclosed AI video of a congressional candidate, and a federal judge ruled it was satire. A federal judge also struck down Montana's similar law because its satire exemption required subjective judgments about speech.

Hier said the PAC plans a Facebook ad campaign using the Hardy video and a new AI video of Sen. Dave Wolk. The incident shows how little a labelling law does on its own when the publisher claims satire. Campaigns and officials need their own record of what each clip shows and how it was made.

What security teams should do this week

This week's cases did not need new attack techniques. They exploited the gap between a convincing piece of media and the moment someone checks it. These steps close that gap:

• Set up likeness and brand monitoring for named executives, clinicians and spokespeople on the platforms your customers use, and assign an owner for repeat uploads.
• Prepare a takedown evidence pack: original URLs, capture times, the file itself and a forensic report, so a complaint to a platform or police does not start from scratch.
• Require a call-back on a known number before any payment, credential reset or account change requested in an urgent voice call, and train staff that a familiar voice proves nothing.
• Flag repeated urgent wire transfers made under phone instruction for a pause and a verification call, especially for older customers.
• For communications and election teams, agree in advance who can analyse a suspect clip and publish a finding within hours.

Where Neural Defend fits

For the AP-HP and Vermont cases, the fit is NeuroDetect. An analyst uploads the authorised video or audio into a case and gets explainability heatmaps and attention views down to regions, frames and audio segments, a plain-language summary linked to the evidence, and a downloadable forensic report that records observations, uncertainty, provenance and model context. The analyst makes the final call, and the report is the kind of record a complaint or a public correction needs.

For voice-driven fraud against a bank or contact centre, NeuroCall (Beta) analyses call audio for synthetic voices. It would not sit on a private family phone, so the Del Mastro case still depends on call-back habits and on payment providers spotting coercion.

Where a deepfake face reaches onboarding or account recovery, NeuroVerify checks the image, video or audio in one request across six signals and returns a unified risk score in under 200 ms p95 on eligible image requests. Its detection models build on ARCAS 1B, and Neural Defend ranked first of 36 systems on the public VendorBench-100 still-image benchmark, where one author also co-authored the ARCAS paper. Neural Defend holds ISO/IEC 27001, ISO/IEC 27701, ISO/IEC 42001 and SOC 2 Type II, with cloud, private cloud, on-premises and air-gapped deployment.