A hacking incident involving Suno, an AI music generator, has produced leaked materials that reportedly show how its models were trained. According to reporting cited by outlets including Variety and The Verge, the exposed data indicates Suno scraped millions of songs and accompanying lyrics from online platforms and websites, including YouTube Music, Deezer, and Genius. The Verge and Variety describe this as a rare look into training practices that Suno has not publicly detailed, including how its training datasets were obtained.
Slashdot, citing 404 Media, adds additional claimed sources mentioned in the leaked code and dataset documentation, including Pond5, Jamendo, Freesound, and certain podcast RSS feeds. The leaked materials reportedly include scraping instructions, notes about dataset scope and ingested clip counts, and references to filtering “non-music.” They also reportedly suggest the use of third-party scraping infrastructure and tools.
Multiple sources also describe allegations and context from prior record-industry disputes, where critics accused Suno of using copyrighted content. Suno maintains that its training used publicly available music files and metadata and characterizes that approach as fair use. In addition to training-related materials, the leak report says the attacker accessed customer information such as emails/phone numbers and certain payment details.
