Suno AI Music Training Hack Exposes Millions of YouTube Songs Used Without Permission

In November 2025, a hacker operating under the alias ellie.191 broke into Suno's systems using a supply chain attack and pulled back the curtain on how the AI music company built its generative models. What they found—and what 404 Media reported in July 2026—was a detailed map of the company's training pipeline: over 2 million music clips scraped from YouTube Music, more than 17,000 hours of audio from Genius HQ, over 12,000 hours from the streaming service Deezer, and more than 62,000 hours from Pond5, a stock music library owned by Shutterstock. None of this material came with permission from the artists, songwriters, or rights holders who created it.

For years, musicians and record labels have accused AI music companies of quietly harvesting their work to train algorithms that could eventually replace them. The lawsuits came first—Universal Music Group, Sony Music Entertainment, and Warner Music Group filed suit against Suno claiming copyright infringement and unauthorized use of their catalogs. The hack provided something those companies had been seeking: concrete evidence of what was actually in Suno's training data and where it came from. The source code files, screenshots, and documentation shared with journalists showed exactly which platforms had been scraped and how much material had been pulled from each one.

The breach also exposed something else: customer payment information held by Suno's payment processor, Stripe. The hacker gained access to records and financial details, though Suno later stated that full credit card numbers were not compromised. The company moved quickly to contain the damage, investigating the breach and concluding that the exposed source code was outdated and no longer in active use. Under current privacy law, Suno determined it had no obligation to notify individual users about the incident.

Suno's response to the revelations was measured. A company spokesperson acknowledged the breach but characterized the leaked code as legacy material—old infrastructure that had been replaced. The company also pushed back on the implications, noting that it had already disclosed its training methods publicly on its website and in regulatory filings. "Suno's music generative AI models are trained on publicly available music files and related metadata accessible on third-party websites on the open internet," the company's statement read. In other words: we took what was there to take.

That framing sits at the heart of the legal battle now unfolding. Suno, like other generative AI companies, is relying on the fair use doctrine—a provision in copyright law that permits limited use of copyrighted material for purposes like research, criticism, and education. Anthropic and Meta both won lawsuits brought by authors last summer using similar arguments. But the music industry's position is different: they argue that Suno is not transforming the material or creating something fundamentally new, but rather using human-created songs as raw material to build a tool that generates cheap imitations—what creators call "AI slop"—that could displace human musicians from paying work.

The tension reflects a deeper fracture in creative industries. Tech companies frame AI as a democratizing force, a tool that makes music creation accessible to anyone with a text prompt. Artists and their representatives see it as theft dressed up in the language of progress. Some rights holders have negotiated licensing deals with AI companies, establishing a framework where compensation flows back to creators. Suno itself has introduced some guardrails—you cannot ask the system to generate a song in the style of Taylor Swift, for instance. But those measures feel incremental against the scale of what was trained into the system without consent.

What the hack revealed, ultimately, is that the black box of AI training is less mysterious than it seemed. The data came from somewhere. It came from YouTube, from music databases, from streaming platforms. It came from the work of people who were never asked and will never be paid. Whether that constitutes fair use or infringement will be decided in court, but the evidence is now public: millions of songs, scraped and fed into an algorithm, to teach a machine how to make music.