Suno's Source Code Leak Reveals Scraping Network Across YouTube, Deezer, and Podcast Feeds

The Breach and What It Exposed

A security incident last November at Suno, the AI music generation platform, has surfaced new details about the infrastructure behind its training pipeline. An unauthorized actor gained access to source code and internal documentation that mapped out the company's data collection methods, including technical specifics on how the startup pulled audio files from major streaming services and podcast RSS feeds.

The attacker used a worm to compromise credentials belonging to a Suno engineer, opening a path to GitHub repositories and cloud storage instances. What emerged from those systems was not just outdated code, but a blueprint of scraping operations that pulled from YouTube Music, Deezer, Genius for lyrics, and stock music catalogs. The leaked material also included proxy configurations apparently designed to mask requests when harvesting YouTube content, including isolated vocal tracks.

Suno acknowledged the breach in a statement, characterizing it as "limited" and noting that the exposed code was no longer in active use. The company said no sensitive personal information was compromised, though hundreds of thousands of customer records containing email addresses and phone numbers were reportedly part of the accessed data. Suno emphasized it does not store full credit card numbers.

Training Data at Scale

The leaked documentation points to a multi-pronged ingestion strategy. Beyond streaming platforms, Suno appears to have tapped podcast RSS feeds at scale, scraping what internal records suggest were hundreds of thousands of episodes. The approach mirrors patterns seen across the generative AI sector, where companies build training corpora by pulling publicly accessible files and metadata from the open web.

Suno has been transparent in court about the scale of its training set. In a 2024 filing related to an ongoing copyright lawsuit from record labels in the United States, the company disclosed that its models were trained on tens of millions of recordings sourced from the internet. It defended the practice as fair use, a legal doctrine that permits limited use of copyrighted material without permission under certain conditions, such as for commentary, research, or transformative purposes.

That lawsuit, filed by a coalition of major labels, has seen one significant defection. Warner Music Group withdrew late last year after striking a licensing agreement with Suno, a move that signals the industry's split strategy of litigation and dealmaking. Other labels remain plaintiffs, and the case continues to test the boundaries of fair use in the context of machine learning.

The Proxy Layer and Platform Evasion

One technical detail in the leaked code stands out: the use of proxy services to route scraping requests. This suggests Suno engineered its data collection to avoid rate limits or detection mechanisms that platforms like YouTube deploy to block automated harvesting. Proxies rotate IP addresses and distribute requests across geographic regions, making large-scale scraping operations harder to trace or throttle.

YouTube's terms of service explicitly prohibit automated data extraction, and the platform has historically pursued legal action against scraping operations. The presence of proxy infrastructure in Suno's codebase indicates the company anticipated pushback and built evasion into its pipeline. The leaked records also reference acapella versions of tracks, which are valuable for training models to separate vocals from instrumentation, a core capability in generative music systems.

Deezer and Genius, the other named sources, offer different data types. Deezer provides high-quality audio streams, while Genius is the go-to repository for crowdsourced lyrics. Together with YouTube, these platforms form a comprehensive training corpus that spans audio fidelity, lyrical content, and genre diversity.

Customer Data in the Wild

The breach extended beyond training infrastructure. Customer records for what the hacker described as hundreds of thousands of Suno users were part of the accessed dataset. These records included email addresses and phone numbers, common identifiers that can be used for phishing or credential stuffing attacks.

Suno's statement noted that the company does not have access to full credit card details, which are handled by Stripe, the payments processor. This architectural choice limits exposure in the event of a breach, though the leaked contact information still poses risks. The company said it determined that individual notifications to affected users were not required under applicable privacy regulations, a threshold that varies by jurisdiction and depends on the sensitivity of exposed data.

The Fair Use Argument Under Pressure

Suno's legal defense rests on the argument that training AI models on copyrighted works is transformative and therefore protected under fair use. The doctrine is well-established in areas like parody, criticism, and search indexing, but its application to generative AI remains contested. Courts have yet to issue definitive rulings, and the Suno case is one of several that will shape precedent.

The fair use test in U.S. law weighs four factors: the purpose and character of the use, the nature of the copyrighted work, the amount used, and the effect on the market for the original. Suno argues its models create new works rather than replicate existing ones, and that training is a non-expressive intermediate step. Rights holders counter that generative AI directly competes with human creators and that scraping entire catalogs exceeds what fair use permits.

Warner's licensing deal complicates the narrative. By opting out of litigation in favor of a commercial arrangement, the label effectively endorsed Suno's technology while securing revenue. Other labels, including Universal Music Group and Sony Music Entertainment, continue to pursue the lawsuit, betting that courts will side with rights holders and establish stricter rules for training data.

Industry-Wide Scraping Practices

Suno is far from alone. A recent investigation identified datasets used by multiple AI music startups that collectively contain millions of copyrighted tracks, scraped from streaming services and online repositories. These datasets circulate in research communities and are used to benchmark models, often without clear legal authorization from rights holders.

The practice has drawn scrutiny from regulators and advocacy groups. The European Union's AI Act includes provisions that require transparency around training data, and ongoing negotiations over copyright exceptions for AI training in the EU and UK will determine whether scraping for commercial purposes remains viable. In the U.S., legislative proposals have stalled, leaving the issue to courts.

For platforms like YouTube and Spotify, the scraping arms race is a persistent challenge. Both invest heavily in anti-bot measures, but determined actors with sufficient resources can circumvent defenses. The proxy infrastructure revealed in Suno's code is standard practice in large-scale data collection, used by academic researchers and commercial actors alike.

What Comes Next for Suno

The breach and subsequent leak place Suno in a difficult position. While the company maintains its training methods are legal, the technical details now in the public domain will likely be scrutinized by plaintiffs in the ongoing lawsuit. Evidence of proxy use and systematic scraping may undermine arguments that Suno's data collection was passive or incidental.

The customer data leak, though described as limited, adds reputational risk. Users of generative AI tools are already wary of how their inputs are stored and used, and a security incident that exposes contact information erodes trust. Suno's decision not to notify users individually may comply with legal minimums, but it sidesteps the broader question of transparency.

Looking ahead, the company faces a choice: continue to defend its scraping practices in court, or follow Warner's lead and pursue more licensing deals with rights holders. The latter path offers legal certainty but comes with costs that could reshape Suno's unit economics. Training on licensed catalogs is expensive, and the company would need to pass those costs on to users or accept lower margins.

The breach also underscores a wider vulnerability in the AI sector. As startups race to build foundation models, security often lags behind product development. Credentials stored in plaintext, outdated code left in repositories, and inadequate access controls are common weaknesses. For Suno, the incident is a reminder that the same infrastructure used to scrape the web can be scraped in turn.