A Hacker Breached Suno and Exposed Exactly How It Trained On YouTube, Deezer, and Genius

Suno’s founder, Mikey Shulman, has said the quiet part out loud before: “most people don’t like making music, they like the feeling of having made something.” His company, Suno, turns a one-line text prompt into a finished song, complete with vocals and instruments, in about a minute. It is one of the most-used AI music tools around, and until this month, exactly what it learned to sing from was mostly a matter of lawsuits and guesswork. On July 15, 2026, 404 Media reported that a hacker going by ellie.191 had breached Suno’s systems and handed over internal source code answering that question directly. The code shows Suno ingesting well over a hundred thousand hours of music scraped from YouTube, Deezer, Genius, and several other platforms, sorted and labeled for training.

How the hacker got in

The break-in did not start with Suno at all. Ellie.191 says the entry point was Shai-Hulud, a self-spreading worm that has been tearing through developer supply chains in 2026 by stealing GitHub publishing tokens and cloud credentials from infected packages, then using them to break into the next target. DecodeStack has covered how that worm spreads in a deep dive on 2026’s software supply chain attacks. In Suno’s case, the worm reportedly compromised an employee’s credentials, and the hacker used that foothold to pull source code dated 2023 and 2024 straight out of the company’s systems.

What the leaked code actually shows

The stolen files were not just code, they carried comments and internal names for each data source Suno’s engineers had built pipelines around: genius_hq, youtube_music, freesound, jamendo, imp (the IMSLP sheet-music archive), deezer, and a bundle called ytm_tagged. The scale attached to each name is what makes the leak more than a rumor confirmed:

• YouTube Music: more than 113,000 hours across over two million individual clips
• ytm_tagged: a separate, labeled YouTube corpus of more than 152,000 hours
• Pond5: more than 62,000 hours from the stock-audio library
• Genius: roughly 17,600 hours, pulled from a site better known for song lyrics than audio
• IMSLP: roughly 19,500 hours from the public sheet-music archive
• Deezer: more than 12,000 hours from the streaming service

404 Media puts the combined total at the equivalent of several decades of continuous music. The code reportedly goes further than a simple download list: it shows Suno specifically targeting a cappella versions of songs on YouTube to isolate vocals for training, and routing the scraping through Bright Data, a commercial web-scraping company, to get around YouTube’s own anti-scraping defenses.

Ellie.191 also claims to have pulled Suno’s customer records: names, emails, phone numbers, and Stripe payment details for a portion of its user base. Suno told 404 Media it identified a “limited security incident” in November 2025, contained quickly, involving outdated source code, and says no sensitive personal data was exposed; the company adds that it never has access to full card numbers through Stripe in the first place.

Why this lands harder than the average leak

Suno was already being sued over exactly this question. The RIAA, the US recording industry’s trade group, had accused the company in court of ripping songs directly off YouTube by working around the technical protections meant to stop that kind of copying, a claim Suno had not confirmed in public. The leaked code does not prove a new allegation, it corroborates one the industry had already made, with the company’s own internal file names and hour-counts attached. Suno has already admitted, in its own court filings, to training on what it called “essentially all music files of reasonable quality that are accessible on the open internet,” tens of millions of recordings, and has defended that practice as fair use, the part of US copyright law that allows using protected work without permission when the result is different enough from the original to justify it. One of the RIAA’s lawsuits against the company has already been settled.

Responding to 404 Media, Suno leaned on a separate defense: that its models are built for “original creation by design.” The company says it deliberately strips artist names out of the data used to train its models, specifically to stop the AI from learning to imitate a given performer’s style, and points to its investment in tools meant to catch impersonation. That claim is about what the model does with the training data, not about where the data came from, and the leak speaks only to the second question.

The pattern this fits into

This is not an isolated incident so much as a recurring shape: an AI company scrapes broadly to build a training set, faces a lawsuit over it, offers a fair-use defense in court, and only much later does the public get a concrete look at what was actually scraped, usually because something broke rather than because the company volunteered it. The RIAA’s accusations sat as allegations for months; a stolen file with real hour-counts changed that overnight. The Shai-Hulud entry point matters here too: a supply chain worm that steals developer credentials does not just threaten code, it can walk straight into whatever the code touches, including the exact data pipelines a company would rather keep private mid-lawsuit.

Where this leaves Suno

A leak does not settle a fair-use case, but it removes the ambiguity a company can otherwise argue from. Suno’s legal defense rests on how a court reads what it did with copyrighted material it does not deny using; the leak makes that starting point a lot harder to contest with the specific platforms and hour-counts now on the record. For the rest of the AI industry, the lesson is not really about music: any company defending broad web scraping in court is one compromised employee credential away from its own internal file names becoming the plaintiff’s exhibit.

Frequently asked questions

Who leaked Suno’s training data?

A hacker using the handle ellie.191 breached Suno’s systems and gave 404 Media internal source code from 2023 and 2024 detailing the company’s training-data pipelines, first reported on July 15, 2026.

How did the hacker get into Suno’s systems?

The hacker says the entry point was Shai-Hulud, a self-spreading worm that steals GitHub tokens and cloud credentials from infected developer packages. It reportedly compromised a Suno employee’s credentials, which the hacker then used to pull source code out of the company’s systems.

What platforms did Suno scrape for training data?

The leaked code names YouTube Music, Genius, Deezer, Pond5, IMSLP, and a labeled YouTube corpus called ytm_tagged, together amounting to well over 100,000 hours of music, the equivalent of several decades of continuous audio.

Is what Suno did illegal?

That is what the RIAA’s lawsuits against Suno are meant to decide, and one has already been settled. Suno has admitted to training on music scraped from the open internet and argues the practice is protected under fair use; the leak corroborates the RIAA’s scraping accusations but does not by itself resolve the legal question.

Was customer data exposed in the Suno hack?

The hacker claims to have also obtained Suno customer names, emails, phone numbers, and partial Stripe payment data. Suno says it identified a contained security incident in November 2025 involving outdated source code and maintains no sensitive personal data was compromised.