The Count
Resemble AI, a cybersecurity firm that spends its working hours cataloguing synthetic media the rest of us would prefer not to think about, published its H1 2026 Deepfake Threat Report this week. The researchers combed through 1,760 news reports, verified 821 distinct deepfake attacks, and tallied 3.46 million synthetic images, videos, and audio files connected to them. Then they went looking for the tool behind the files. In the report's own words: "one tool accounts for 87% of every file we can count." That tool is Grok, the image generator built by xAI.
Even now, "one tool" is doing a lot of diplomatic lifting for a sentence that could just as easily read "one company."
One in Six
And yet — as if this were not enough — the report's darker number isn't the market share. It's the content. Of the 821 attacks Resemble AI verified, 137 — roughly one in six — involved nonconsensual sexual imagery, and that count includes both adults and children. The firm puts the statutory civil exposure from the documented cases at $2.24 billion, which is the sort of figure that sounds like a typo until you remember it's describing real people's photographs turned into abuse material without their knowledge or consent.
Fifteen thousand seven hundred and thirty-six victims is not a rounding error. It is a phone book.
What xAI Has Said About This
As of this report, nothing on the record specifically addresses the 87% figure. xAI has previously stated it prohibits nonconsensual explicit imagery in Grok's usage policy, a claim that continues to coexist peacefully with Grok topping cybersecurity firms' deepfake-attribution charts report after report. Sources within the Trust and Safety Community note that a usage policy is not the same thing as an enforcement mechanism, in the same way a "no smoking" sign is not the same thing as a working smoke detector.
The Algorithm was, once again, unavailable for comment. It rarely has to be — the report speaks for it.
Who Is Actually Doing Something About It
Resemble AI's pitch, naturally, is that better detection tools — the kind it happens to sell — could close some of the gap between attack and attribution. That's a real product doing real work, and also, worth noting, a company selling smoke detectors publishing a report about how much smoke there is. Both things can be true. What can't be argued with is the underlying pattern: a single consumer AI product now accounts for the overwhelming majority of traceable deepfake harm in a six-month window, and the response from its maker has so far been silence dressed up as a policy document.
The industry keeps promising safeguards are coming. The victims, in the meantime, keep arriving first.
Sources: Resemble AI — H1 2026 Deepfake Threat Report · Tech Times — Grok Deepfakes Hit 15,736 Victims in Six Months
