Hacked source code reveals how Suno scraped decades worth of music and podcasts from the internet to train its AI tool.

The code also suggested that Suno was using proxies to scrape songs from YouTube through a company called Bright Data, which sells scraping tools, infrastructure, and data services. Additional code shows that with the help of an online tool called PodcastIndex, Suno identified 420,000 different podcasts that had at least five, 30-minute episodes and sought to download roughly 1 million hours of podcasts

The hacker, ellie.191, told 404 Media they breached the company by hacking an individual employee using the Shai-Hulud worm, a supply chain attack that allowed hackers to harvest GitHub and cloud service credentials. They said they also accessed Suno’s customer list, which included customers’ emails and/or phone numbers and Stripe payment details, depending on what they used to login. The hacker provided a sample of some of the customers, some of whom confirmed to 404 Media they had used their phone number to sign up for Suno and said they were never notified of a breach.

Last month, The Atlantic reported on several music databases that are widely used in AI training, consisting of millions of tracks: “Three of the datasets I found are distributed as a list of links to songs on YouTube or Spotify. AI developers download the actual audio using tools that automate the job, some of which allow developers to bypass logins, advertisements, and mechanisms that might earn money or subscribers for creators. Such tools violate the terms of service of these platforms.

Archive link: https://archive.ph/xX3XW

  • Mangoholic@lemmy.mlEnglish
    4·
    1 day ago

    If music generated gets diffused out of training data, you could have a map of percantages which songs are used in the generation? We could then pay artist based on this percentage. They would also need to opt in ofc for this to be moral.

    • Jrockwar@feddit.ukEnglish
      1·
      20 hours ago

      Unfortunately AI doesn’t work like that. Any way to explain it would be an oversimplification but I can try.

      The training data (songs) are used to create the weights. This is a bunch of numbers that are on their own meaningless - they don’t map to specific songs, but to attributes such as “tone” “rhythm”… And like that but many (millions of) abstract attributes that don’t make sense as people, but make sense to computers.

      So if the thing makes a song that is very “rhythmic” but also “tonal”, there’s no specific training song that contributed to that - all did, and it’s a mess to decypher how much each contributed to each attribute. Except the resulting song doesn’t use two parameters, uses many millions, so it’s essentially impossible to know.