Suno Got Hacked — and the Leaked Source Code Shows Exactly How It Trained on Millions of Copyrighted Songs

In November 2025, someone broke into Suno. Not through a zero-day exploit or a misconfigured firewall — just a supply chain attack that compromised employee credentials. The hacker grabbed source code and internal records, then handed the whole thing to 404 Media. Eight months later, in July 2026, the public finally learned what was in those files.

It’s worse than anyone in the AI music business wanted to admit.

It’s not a training data controversy. It’s evidence.

The leaked materials, reviewed by 404 Media’s Jason Koebler, show that Suno trained its AI models by scraping decades of audio from YouTube Music, Deezer, Genius, multiple stock music libraries, and podcast RSS feeds. Millions of songs. Millions of lyrics. All fed into the training pipeline without asking.

Suno has never been transparent about its training data. When pressed, the company’s line has been consistent: it trains on “publicly available music files” on the open internet. In the RIAA’s ongoing lawsuit, Suno went further — openly admitting it trains on copyrighted material and arguing that doing so constitutes fair use.

The leaked source code draws a line through that argument.

YouTube’s terms of service explicitly forbid automated scraping. The Digital Millennium Copyright Act prohibits circumventing platform protections against data collection. Suno did both, according to the internal records. Its competitor Udio has been accused of the same thing.

“Fair use” is a legal defense, not a license. And when you have to bypass technical barriers — when you have to route around YouTube’s anti-scraping measures — you’re not operating in a gray area anymore. You’re choosing to break rules and hoping the fair use argument covers it later.

There’s an irony here worth noting. Genius, the lyrics database that Suno scraped, built its own catalog in part by scraping the Original Hip-Hop Lyrics Archive, a Web 1.0-era database. The scraper got scraped. The difference now is scale: Suno didn’t grab a few thousand text files. It pulled millions of songs across multiple platforms, industrializing what Genius did manually two decades ago.

The RIAA lawsuit isn’t just about one company. It’s a test case for whether copyright law still functions when training data is measured in petabytes. If Suno loses, the entire AI music industry has to rebuild its training pipelines from scratch — or license everything, which would fundamentally change the economics. If Suno wins, “fair use” expands into territory that Congress never envisioned when it wrote the doctrine. Either outcome sends shockwaves through AI image, video, and text generation too.

The breach Suno forgot to mention

The training data revelation is the headline, but the breach itself deserves scrutiny — especially given how Suno handled it.

The hacker didn’t just walk away with training pipeline details. Customer data was in those files too: email addresses, phone numbers, and partial credit card numbers through Stripe. Suno never notified affected users. Not a single email. Not a blog post. Not a mandatory regulatory filing that anyone can find.

Here’s the company’s official statement to Pitchfork, delivered after 404 Media broke the story: “In November of 2025, we determined that Suno had been the subject of a limited security incident that was quickly contained. At the time, we immediately conducted an investigation and verified that the incident primarily involved outdated source code that is no longer in use at Suno and that no sensitive personal information was compromised.”

Read it twice. Notice the sleight of hand. “Outdated source code that is no longer in use” — this frames the training data pipeline as legacy, as if the scraping happened in some earlier era of the company rather than being foundational to how their product works. “No sensitive personal information was compromised” — but partial credit card numbers, emails, and phone numbers were in the leaked files. Ask anyone in infosec whether those count as sensitive.

The company also emphasized that it “does not have access to customers’ full credit card numbers in Stripe.” That’s a precise wording. It tells you what they don’t have. It doesn’t tell you what the hacker did get.

Eight months. A “quickly contained” incident. Customers learned about it from a news article — not from the company that lost their data.

The transparency black hole

Suno is a case study, not an outlier. The entire generative AI industry — text, image, music — operates behind the same curtain. OpenAI stopped disclosing GPT-4’s training data details. Anthropic won’t say what Claude was trained on. Midjourney is fighting class-action lawsuits from artists, arguing that training data is a trade secret.

What makes Suno different is that we finally have evidence. Not speculation, not inference — actual source code and database records pulled from internal servers by someone who had no ideological axe to grind. The hacker told 404 Media: “I like to hack anything and everything.”

That almost makes it worse. The most revealing glimpse into how an AI company actually builds its models came not from regulation, not from journalism, not from voluntary disclosure — but from a bored hacker.

For musicians, the stakes are existential. When an AI can generate a song in your style in seconds, trained on decades of work you created and own, “fair use” starts to sound less like a legal doctrine and more like a punchline. The Human Artistry Campaign, a coalition of music industry organizations, has been protesting exactly this scenario. Their argument is simple: you can’t take creators’ work to train AI, then use AI-generated content to compete against those same creators. That’s not innovation — it’s extraction.

The RIAA lawsuit is the mechanism that might force an answer. It’s not about whether AI can learn from public data. It’s about whether “publicly available” means the same thing as “free to industrialize,” and whether breaking platform rules to get the data invalidates any fair use claim you might have had.

Google’s position in all of this is awkward. As YouTube’s parent company, it’s the platform Suno allegedly scraped. But Google is simultaneously fighting copyright lawsuits from major book publishers who accuse it of scanning their works to train AI models. One arm of the company is the victim in the Suno story. Another arm is the defendant in an almost identical case. The cognitive dissonance is hard to miss.

There’s also the hacker, who told 404 Media they had no particular motivation. “I like to hack anything and everything.” That’s what produced the most revealing document the AI music industry has ever seen. Not a congressional investigation. Not a whistleblower. Not a transparency report. A bored person with technical skills and no agenda.

Think about what that means. The only reason we know how one of the biggest AI music companies actually built its models is that someone hacked them for fun. No regulator demanded disclosure. No journalist uncovered it through reporting. The system that’s supposed to produce transparency around AI training data — voluntary disclosures, terms of service, corporate responsibility statements — produced exactly nothing. A random supply chain attack produced everything.

The precedent problem

The Suno leak matters beyond the immediate scandal. It establishes a baseline of distrust that the entire AI training data debate has been missing.

Until now, the conversation has been theoretical. Companies say they train on “publicly available data.” Critics say they train on copyrighted material without permission. Both sides argue past each other because nobody outside the companies actually knows what’s in the training sets. The Suno hack changes that. We now have a confirmed case — not alleged, not inferred — where an AI company systematically bypassed platform protections to harvest copyrighted content at industrial scale.

That precedent will echo. Every AI company that has used similar language about its training data now faces the same suspicion. “Publicly available” is no longer a safe answer. The burden of proof has shifted. If you can’t show what’s in your training data, people will assume it looks like Suno’s.

The practical question is whether this changes anything about how AI companies operate. Probably not voluntarily. The economic incentives to scrape first and litigate later are too strong. A startup that asks for permission while competitors don’t will get out-trained, out-funded, and out-competed. The only force that can change behavior is legal liability — and that depends on how the RIAA case resolves. In the meantime, the safest assumption for anyone creating music, art, or writing online is that an AI company is probably training on it right now.

Eight months. One hack. A pile of source code that laid bare an industry’s worst-kept secret.

Suno isn’t the only company scraping where it shouldn’t. Udio faces similar scrutiny. Every AI music startup built before 2026 almost certainly trained on data it didn’t have rights to. The question now is whether getting hacked and publicly exposed — source code, customer records, and all — can force this industry into a real conversation about training data transparency. If this level of exposure doesn’t move the needle, nothing will.