Loading today’s stories
Security1 distinct publisher3 min readPublished
The complaint filed Wednesday in Northern California rests on hash values from an abuse series that has circulated since the early 2000s, which puts the contested question on what xAI ingested rather than on what its prompt filter refuses.
The Watch · Security desk
Compiled by The WatchSomething wrong?How this is made
Hash values are the mechanism worth reading twice. The complaint says the plaintiff's abuse series carries well-known hash values, and that those values have surfaced in Grok output spreading on X [5]. It further asserts that CSAM depicting her has been found on xAI through investigative reporting, takedown requests, and criminal cases [6], and that Grok generated images depicting her and the series in which she is the victim [7]. Ingestion and output match are separate questions. A hash comparison speaks to the second, and the plaintiffs are using it to argue backwards to the first.
The ingestion path is pleaded directly. Grok's terms of service treat anything posted on X as training material, so the suit argues any CSAM posted to the platform over the past year was likely absorbed by the model [14]. Her material has been online since at least the early 2000s, with hundreds of thousands of related files in law enforcement submissions to NCMEC [3][4].
The filter, as the complaint describes it, rejects requests only where Grok detects clear intent in the phrasing of the request; indirect or euphemistic prompts slip past a text-based filter, and while the underlying capability remains, some volume of CSAM becomes effectively inevitable [13]. xAI has said it built guardrails against sexualized deepfakes; the plaintiffs call them very weak, divergent from industry practice, and easily circumvented [12]. The suit also contrasts competitors that suppressed nudification with xAI, which it says embedded Grok and its deepfake capability directly into X [11].
Run the only public counts. More than 3 million sexualized images over 11 days is about 273,000 a day [19]. The at-least 23,000 that appeared to depict children is about 2,090 a day [20], or 0.77 percent of the total [21], and those are the figures from the Center for Countering Digital Hate, not from any regulator or from xAI. That fraction is the part a procurement file has to hold: apparent child imagery is a rounding error by volume and the whole of the legal exposure by value.
xAI did not return CyberScoop's request for comment [16]. An earlier suit this year alleges a man generated thousands of deepfake CSAM images of his stepdaughter with Grok, and that xAI withheld information that would have helped identify him; he died by suicide days after investigators traced and seized the material [17]. In July, Musk sued Minnesota Attorney General Keith Ellison over a state law banning nudification technology that carries a $500,000 fine [18].
The artifact this filing implies for buyers is a corpus statement: what was ingested, what was screened out, and who attests to it under subpoena. Refusal testing measures the last layer of the stack and leaves no record of the first.
Ranked by verification strength, evidence, and original report placement.
A class action lawsuit filed Wednesday in the U.S. District Court for the Northern District of California accuses xAI of training Grok's synthetic deepfake "nudify" capabilities on real images and videos of child abuse.
The suit names Jane Doe 1 and other anonymous individuals as plaintiffs and calls Doe an identified victim of child pornography tracked by the FBI's Child Exploitation Notification Program, under which she still receives updates when images related to her abuse surface online.
Doe was pre-school aged when her perpetrator's abuse began, it continued for years for the explicit purpose of making CSAM to distribute online, and media depicting her has circulated online since at least the early 2000s.
The lawsuit states that hundreds of thousands of files related to the victim have been included in law enforcement submissions to the National Center for Missing & Exploited Children.
The suit cites Masha's Law, passed in 2018, which preserves civil legal remedies for victims of child pornography and online exploitation, as the basis for its allegations.
An analysis by the Center for Countering Digital Hate found that during an 11-day period between December 2025 and January 2026, Grok created more than 3 million sexualized images, at least 23,000 of which appeared to depict children.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 27, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
invest
Minnesota asks a federal judge to treat Grok Imagine as a tool, not a speaker3 distinct publishers
product
Grok CSAM suit gains a fourth plaintiff and a 7,000-image count2 distinct publishers
build
Grok Build's real product is the X timeline, not the code generator1 distinct publisher
science
A school laptop, a chatbot, and a principal who said not to worry1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Thin outside the pleading
The documentary layer is solid: one trade outlet reports the filing, venue, statutory basis, requested relief, and a third-party image count. But the load-bearing factual claims, that known CSAM hash values appear in Grok outputs, that Grok generated images of this plaintiff, and that the nudify capability was trained on real abuse material, rest entirely on an untested complaint with no forensic detail, no adjudication, and no response from xAI.
Capability live and heavily used
Adoption of the capability at issue is well evidenced even where the allegations are not: the image feature is embedded in X, and a third-party count places output at more than 3 million sexualized images in 11 days, roughly 273,000 per day, with about 0.77 percent appearing to depict children. This measures usage scale, not the truth of the training-data claim.
Causal claim runs ahead of proof
The headline proposition, that Grok's nudify capability was trained on real CSAM, is stronger than the evidence offered for it, which is a hash-value inference plus a terms-of-service ingestion argument. The gap is moderate rather than large because the adjacent facts are well grounded: the filing is real, the CCDH counts are specific, the prior suit and the Minnesota litigation are documented, and the article attributes allegations as allegations. On the other side, Musk's blanket denial understates what the same reporting documents.
Heavily incentivized on all sides
Every participant has a stake in the framing. Plaintiffs and their counsel seek monetary damages and injunctive relief under a statute built for civil recovery, and the class is pleaded at thousands of members. CCDH is an advocacy organization whose counts advance a platform-accountability agenda. Musk and xAI are simultaneously defending multiple suits and prosecuting an affirmative First Amendment challenge to a state nudification ban carrying $500,000 per-instance penalties, and xAI declined to comment.
Low-moderate
One publisher, one source document, no defense response, and no adjudication. Confidence is adequate for the procedural facts, that this suit exists, what it pleads, what relief it seeks, and for the reported CCDH figures, but low for the substantive training-data and hash-match questions the story turns on.