Leaks Reveal Suno Fed Thousands of Hours of Deezer, YouTube and Pond5 Data Into Its AI

A hacker broke into AI music platform Suno and walked out with supply code that paperwork, in exact element, precisely the place the corporate’s coaching information got here from.

The breach was first reported by 404 Media, which reviewed the leaked recordsdata. It confirms what the music business had been saying in courts since 2024.

The intruder claims to have used a chunk of malware known as the Shai-Hulud worm—named after the big sandworms in Frank Herbert’s Dune. Suno, one of many largest AI music mills on-line, lets customers sort a textual content description and obtain a full tune in seconds; constructing that functionality required a considerable coaching dataset—a set of audio recordsdata used to show the mannequin what totally different genres and types sound like.



The leaked materials consists of scraping directions and inside logs from 2023 and 2024, providing a uncommon have a look at how these pipelines are literally assembled.

The dataset breakdown is restricted. In response to inside file feedback reviewed by 404 Media, the coaching library included 113,879 hours of YouTube Music, 152,162 hours of tagged YouTube tracks, 62,117 hours from inventory music library Pond5, 12,287 hours from Deezer, and 17,615 hours in a dataset labeled genius_hq, related to materials collected by way of Genius. The code additionally documented plans to obtain roughly 1 million hours of podcast audio through RSS feeds.

One inside file monitoring YouTube Music ingestion alone logged 2,013,545 music clips. That is thousands and thousands of recordings overlaying many years of audio—and the urge for food wasn’t restricted to music.

The hacker claimed to have accessed data related to a whole bunch of hundreds of consumers, together with emails, cellphone numbers, and Stripe-related data. Suno disputes that delicate private data was compromised.

The corporate says it recognized the incident in November 2025 and known as it “restricted.” Suno decided the publicity primarily concerned outdated supply code not in use and concluded that particular person buyer notifications weren’t required below relevant privateness legal guidelines. Customers are solely discovering that out now, by way of information protection.

Here is the factor: Suno had already informed anybody prepared to learn its personal web site that one thing like this was occurring. Beneath California’s AB 2013 regulation—which requires AI corporations to reveal their coaching practices—the corporate publicly acknowledged that its coaching information might embrace music “topic to mental property safety,” and listed the corpus at tens of thousands and thousands of publicly out there music audio recordsdata. What the hack provides is specificity: The authorized submitting was obscure by design, and the leaked code will not be.

The scope of AI music coaching was already turning into clear earlier than anybody breached something. In June 2026, The Atlantic printed 4 searchable databases documenting music used to coach AI fashions—one containing 12 million tracks, one other with 9 million, and two extra with round 100,000 every. You possibly can lookup your favourite artist earlier than a hacker handed anybody supply code.

The Recording Trade Affiliation of America had alleged in a 2025 modification to its unique 2024 lawsuit towards Suno that the corporate was ripping songs instantly from YouTube—an accusation Suno contested below a good use protection. The go well with sought $150,000 per infringement incident. The hacked supply code corroborates the RIAA’s central allegation.

Udio, which was focused in a parallel lawsuit filed by the identical major-label coalition, settled with Warner Music in November 2025 and is now transitioning to a licensed platform. Suno’s case with Sony and UMG stays lively in federal court docket; the corporate’s valuation sits at $5.4 billion with round 100 million customers on the platform.

Suno didn’t instantly reply to a request for remark by Decrypt.