The Deepfake Watchlist is Resemble AI's weekly surveillance of synthetic media incidents, ongoing cases, and disputed content shaping the news cycle. Each week we track confirmed incidents, emerging attack vectors, and claims under investigation, alongside the provenance, detection, and policy threads running underneath them. New to deepfakes? Start with our Deepfake 101 Guide.
★ Featured: A network of 293 accounts stole real creators' faces to feed Canadians fabricated trade-war news
McGill University's Centre for Media, Technology and Democracy published Stolen Faces: An Inauthentic TikTok and Instagram Network Inflaming the U.S.-Canada Trade War documenting at least 293 accounts that put AI voices and redrawn lips on real influencers' footage, and CTV News covered the findings on November 16.
- Category: Political / Electoral
- Type: Attack
- Modality: Video, Audio
- Policy / Regulatory: The Canada Elections Act covers impersonation in an electoral context only, and the researchers note that the proposed Safe Social Media Act (Bill C-34) would require labels on synthetic content but would do little about impersonation or coordinated inauthentic activity.
- Trend: Influence networks are building on stolen footage of real creators instead of fully synthetic avatars, which keeps their videos outside what platform AI labels currently catch.
- Attack vector: Footage lifted from creators' own videos is re-voiced with AI text-to-speech and lip-synced to new scripts, many of them AI-written, then posted at volume from accounts the researchers trace to East or Southeast Asia.
- What the researchers documented: The report lays out the technique with measurements readers can check against its side-by-side videos, for example:
- Head, hair and eyes in the fakes match the creators' original frames closely, while the mouth does not, because the lips were redrawn to fit a new voiceover.
- One account posted six reels built on the same clip of creator Andrei Jikh, with identical head movement and six different scripts, one saying Canada's economy was collapsing and another saying Canada was building a future beyond the United States.
- An open-source speech deepfake detector rated the narration as AI-generated in 94% of the network's 3,201 videos, which drew 22.1 million plays between August 7 and November 4.
- TikTok's AI label caught 22 of the 24 videos that used fully AI-generated avatars, but only 26 of the 709 videos built on real people's footage, and none of the network's videos disclosed AI use.
TikTok was fairly good at spotting a face that a model invented from scratch, and almost blind to a real face with a synthetic mouth and a synthetic voice. Only 6.3% of commenters said or hinted that a video might be fake, while one in ten spoke to the narrator as a real person.
Much of the content was simply false. The network's most-played story, a "48-hour ultimatum" from Trump that Prime Minister Carney supposedly answered by cancelling $8 billion in U.S. defence contracts, never happened, and it still ran across 161 videos for 5.1 million plays. The network argued both sides of the trade fight, attacking Carney's government in some videos and the U.S. administration in others, and the researchers say they do not know who pays for it or why. All 293 accounts were still active as of November 11.
The signal that held up here was the audio. A visual check asks whether a face looks wrong, and these faces were real. A check on the voice track asks whether the speech was synthesized, and that question still has an answer when everything around it is borrowed. The researchers identified 104 of the 226 people whose faces were taken, and most of them likely had no idea. We presume TikTok and Meta will act on the accounts now that the report is public, and the next step is checking the voice and lip movement when a video is uploaded, so a network like this gets caught at its first post.
1. The FBI took down North Korea's "FaceStealer" scheme after deepfake video calls helped steal more than $400 million
The Justice Department announced that the FBI had seized the cloud infrastructure behind FaceStealer, a North Korean scheme that used recorded webcam video and deepfakes in fake video meetings to steal more than $400 million in cryptocurrency over 18 months, and The Hacker News and The Korea Times reported on the scam.,
- Category: Fraud / Impersonation
- Type: Response
- Modality: Video
- Policy / Regulatory: The FBI seized the infrastructure and more than 900 webcam videos of victims under a federal warrant, published a technical advisory through IC3, and set up a resource page at fbi.gov/facestealer.
- Trend: State-backed hackers are turning each compromised victim into bait for the next one by replaying that person's own recorded face.
- Attack vector: Meeting invites arrive from the hijacked accounts of earlier victims, prerecorded webcam footage or deepfakes of trusted crypto figures play inside the call, and a fake audio error prompts a "fix" that installs malware.
The part of this scheme that should change how people treat video calls is the reuse. DOJ says the hackers recorded each new target's webcam during the meeting, then played that footage to the target's own contacts in later meetings. Every victim became a credible face for the next call, which is why the FBI's Los Angeles office called it capable of deceiving even the most security-conscious users.
The deepfakes of well-known crypto figures sat on top of that loop, and the malware did the stealing. DOJ says the group has also started aiming the technique at open-source maintainers to reach the software repositories they control, which turns a crypto theft operation into a possible supply-chain attack. The seizure interrupts the loop for now, and the FBI's advisory covers protective steps and what to do after a compromise. Whatever the advisory says, a link sent mid-call to fix an audio problem is exactly the moment to stop and check who is on the other end.
2. Paris Hilton returned to Capitol Hill to push the House on the stalled DEFIANCE Act
CBS News' Paris Hilton returns to Capitol to push for House vote on deepfake porn bill reports that Hilton joined the bill's cosponsors, Democratic Rep. Alexandria Ocasio-Cortez and Republican Rep. Laurel Lee, on November 17 to press House leadership for a vote on the DEFIANCE Act, which the Senate passed by unanimous consent in January.
- Category: CSAM / NCII
- Type: Response
- Modality: Image, Video
- Policy / Regulatory: The bill would let victims sue people who create or distribute sexually explicit deepfakes of them, recover profits made from the images, and file within 10 years of discovering the violation or turning 18, building on the criminal penalties and 48-hour removal rule in the TAKE IT DOWN Act.
- Trend: Survivors and lawmakers from both parties are pushing for a civil remedy to sit alongside the federal criminal law that already exists.
- Attack vector: Not applicable because this is a legislative response story.
The DEFIANCE Act fills a specific hole. TAKE IT DOWN made posting nonconsensual intimate imagery a federal crime and required platforms to remove it within 48 hours of a request, but a prosecution depends on someone else deciding to bring the case. A civil right of action lets the person in the images go to court on their own terms, which is the shift Ocasio-Cortez described as returning power to survivors.
Hilton said her team found sexual deepfakes online of 83 of the 156 women serving in the current Congress, which would mean more than half of the women voting on this bill have been targeted themselves. The bill has now passed the Senate twice, in 2024 and again this January, and CBS reports that Speaker Mike Johnson's office did not respond when asked why it has stalled. With this Congress ending in January, the next few weeks decide whether it reaches the floor or has to start over.
3. Coinbase says its AI stopped $4.7 million from reaching a deepfake Musk livestream scam
BleepingComputer's Coinbase blocks $4.7 million in transfers to Elon Musk deepfake scam reports that Coinbase says its fraud-prevention system blocked about 91% of the transfers customers tried to send to a fake YouTube livestream of Tesla's October "flying car" demo, where a deepfake Musk promised to double any crypto sent to an on-screen wallet.
- Category: Fraud / Impersonation
- Type: Attack
- Modality: Video
- Policy / Regulatory: Coinbase disclosed the figures in a November 17 blog post, and CoinCentral notes that the company did not say how much reached the scammers or provide an outside audit of the $4.7 million figure.
- Trend: Deepfake giveaway scams are timed to live product launches, when viewers expect the real executive to be on screen.
- Attack vector: A fake YouTube livestream styled as Tesla's official broadcast showed a deepfake Musk, wallet addresses and a QR code, and drew more than 500,000 viewers.
Musk giveaway fraud has run on hijacked accounts and doctored clips for years. What the deepfake adds is a live, talking Musk at the exact moment a real Tesla event is streaming, so the fake arrives when the audience is primed to believe it. Coinbase says hundreds of its customers tried to send millions within about two hours, and that its system had never seen this variation of the scam before.
The 9% that got through matters as much as the 91%. CryptoSlate describes one customer who saw Coinbase's warning, came back, and sent the payment anyway. A check at the transaction can stop money that a person has already decided to send, and the decision itself was made while that person watched a convincing fake. Each step earlier that detection moves, toward the stream itself, leaves less for the last line of defense to catch.
Honorable mentions
A Warsaw court ruled for a narrator whose voice was cloned for an ad. On November 13, the Warsaw District Court found that JFC Polska violated narrator Jarosław Łukomski's personal rights by using a synthetic voice deceptively similar to his in ads for its products, Rzeczpospolita reported. The court ordered the company to remove the ads, apologize and pay more than 47,000 zł, in what Polish media describe as the first ruling of its kind in the country.
Insurers and Visa put numbers on AI-enabled fraud for Fraud Awareness Week. The National Insurance Crime Bureau's 2026 threat assessment, released November 16, found that mentions of AI in questionable insurance claims rose more than 1,400% from 2025 to mid-2026, Insurance Business reported. The figure counts claims that mention AI, and NICB does not present it as a count of confirmed AI fraud. Visa's Fall 2026 Biannual Threats Report followed on November 17, identifying more than $1.4 billion in potential scam fraud in the first half of 2026, up 40% from the prior six months, and listing deepfakes among the tools scam networks use to reach victims across languages.
The pattern

- Three of this week's stories turn on the same move, which is borrowing a face people already trust. The McGill network took real creators' footage, FaceStealer replayed earlier victims' webcam video to their own contacts, and the Coinbase scam put a deepfake Musk on a stream styled as Tesla's real broadcast. In each case the face was familiar for a good reason, and the deception rode on what was attached to it: a synthetic voice, a malicious link or a wallet address.
- The defenses that worked this week arrived too late. The FBI seized FaceStealer's infrastructure after $400 million was already gone, Coinbase stopped transfers after customers had decided to send them, and TikTok's labels missed nearly every video built on a stolen face. Each of those is real progress, and each one shows how much easier the job gets when detection runs on the voice track and the lip movement at the moment content is posted or a caller joins.
- The law is catching up from the victim's side. The DEFIANCE Act would give people targeted by sexual deepfakes a way to sue, and the Warsaw ruling treats a cloned voice as part of a person's protected rights. The McGill researchers note that Canadian privacy torts give the creators in their report limited individual recourse and nothing that addresses the wider harm, which makes impersonation protection a reasonable next item for Ottawa.
Watching next week
- The DEFIANCE Act in the House. Whether House leadership schedules a vote before this Congress ends in January, after the bill cleared the Senate for a second time.
- Platform action on the McGill network. All 293 accounts were still active as of November 11, so the test is whether TikTok and Meta remove them and explain how the videos slipped past their AI labels.
- FaceStealer follow-up. Whether the 900 seized videos lead to victim notifications or charges, and whether more open-source maintainers report being targeted.
- EU AI Act marking deadline. Providers have until early December to make AI-generated content machine-readable under the AI Act's transparency rules, RTÉ reported, and the first signal will be which platforms say they are ready.
The Deepfake Watchlist publishes every Thursday. Subscribe to receive it in your inbox, or follow Zohaib Ahmed on LinkedIn for the weekly social companion. Track every documented incident in the Resemble Deepfake Incident Database, and read the full methodology in our 2026 Midyear Deepfake Threat Report.

