A data breach exposed the actual scraping instructions behind Suno's AI music model — including over two million YouTube Music clips and a third-party scraper called Bright Data — turning a fair-use legal argument into a data-provenance receipt.
Every AI training-data dispute eventually arrives at the same demand: show your work. Frontier labs training on text have mostly dodged that reckoning by settling or litigating into ambiguity. Suno, the AI music generator, has just had the work shown for it. Hacked internal files obtained by a person identifying as “ellie.191” and reported by 404 Media, via The Verge’s Jess Weatherbed, detail exactly how Suno assembled its training corpus — scraping instructions targeting YouTube Music, Deezer, Genius, Pond5, Jamendo, Freesound, and the International Music Score Library Project, plus a third-party scraping vendor called Bright Data allegedly used to pull audio off YouTube. One file shows Suno had ingested 2,013,545 YouTube Music clips by the time it was last updated — a hard number, in a market where training-data provenance is almost never quantified this precisely.
the receipts, itemized
The leak reads like an invoice for data nobody paid for.
Beyond the YouTube Music tally, another file cited by 404 Media and reported by The Verge shows Suno’s datasets included hundreds of thousands of hours of YouTube Music, thousands of hours each from Deezer, Genius, IMSLP, Jamendo, and Pond5, and hundreds of hours of Freesound and MuseScore lyrics. Leaked code also reportedly shows Suno searching YouTube for a cappella versions of songs specifically to source clean vocal-only audio — a targeting choice that undercuts any claim the scraping was incidental. Suno was also apparently reaching beyond music entirely, seeking roughly one million hours of podcasts through a tool called PodcastIndex, according to additional leaked code. Asked about the findings, an unnamed Suno spokesperson told 404 Media only that “Suno’s AI models have been trained on publicly available music files and related metadata accessible on third-party websites on the open Internet” — the same framing the company has used in litigation, now sitting awkwardly next to scraping code that names specific protected platforms.
a defense built on definitions
Suno has never denied training on copyrighted material — it has argued the law doesn’t care.
In the RIAA’s 2024 suit — filed on behalf of Sony Music Entertainment, UMG Recordings, and Warner Records, according to the RIAA — Suno CEO Mikey Schulman told Mashable his models train on “medium- and high-quality music we can find on the open internet,” adding that “much of the open internet indeed contains copyrighted materials, and some of it is owned by major record labels.” RIAA chief legal officer Ken Doroshow countered at the time that “Suno and Udio are attempting to hide the full scope of their infringement rather than putting their services on a sound and lawful footing.” An amendment to that suit already alleged Suno unlawfully “stream ripped” tracks off YouTube — an allegation the leaked scraping code appears to corroborate, per The Verge’s reporting.
outputs, not inputs
A second, newer legal front tries to route around the fair-use fight entirely.
An independent-artist class action filed by country musician Tony Justice and his 5th Wheel Records, covered by Music Business Worldwide, doesn’t just argue the training was illegal — it argues specific Suno outputs substantially copy specific recordings. Suno’s August 18 motion to dismiss leans on Section 114(b) of the Copyright Act, arguing that “no Suno output contains anything like a ‘sample’ from a recording in the training set, so no Suno output can infringe the rights in anything in the training set, as a matter of law.” The filing cites the June 2025 Bartz v. Anthropic and Kadrey v. Meta rulings as fair-use precedent for training itself, per MBW. Justice’s suit, meanwhile, argues the platforms “undermine the existing streams of revenue, including licensing markets, for independent music,” according to MusicTech’s reporting — the same licensing markets Suno’s scraping code appears designed to bypass.
the price of the data
None of this is happening in a vacuum where the labels are only suing.
MusicTech has reported that the same major labels behind the 2024 RIAA suit are now discussing licensing deals with Suno and Udio that would include fees and a small equity stake in the companies — meaning the market is simultaneously litigating the price of unlicensed scraping and negotiating the price of licensed access, with the leaked training data now serving as leverage for the latter. Separately, the breach exposed Suno customer emails, phone numbers, and Stripe payment details; Suno told 404 Media it discovered the incident in November 2025, that it “primarily involved outdated source code that is no longer in use,” and that it determined “individual notifications were not warranted under applicable privacy laws” — a judgment call some affected customers, per 404 Media’s reporting, learned about only from the leak itself. Whether courts credit Suno’s outputs-not-inputs defense will decide a lot, but the leaked scraping logs have already done something litigation hasn’t: put a real number, 2,013,545 clips, on what “publicly available” has actually meant for one of AI music’s flagship models. Watch whether that number moves the label-side licensing talks toward a price, or just hardens both sides for trial.
A file for YouTube Music notes that Suno had consumed 2,013,545 YouTube Music clips at the point it was last updated.
