On July 15, 2026, the internal source code of Suno — one of the largest AI-based music generators — was leaked to the public. This leak achieved what plaintiffs had been trying to prove in court for years: it provided precise figures on where the service sourced its training data. The answer — from all over the internet, through industrial scraping, on a scale that is physically impossible without a proxy infrastructure. Let's explore what exactly was uncovered and why this story is about the foundational layer of the entire AI industry, not just one service.
What Happened: The Hack Exposed Suno's Operations
According to 404 Media and TechCrunch, a hacker using the alias ellie.191 gained access to Suno's systems back in November 2025 through a supply chain attack — the Shai-Hulud worm, which collects GitHub and cloud service credentials. By compromising an employee's account, the attacker downloaded the source code and internal configurations, along with information about hundreds of thousands of Suno's clients and payment data from Stripe (emails, phone numbers, partial card numbers).
The incident itself occurred at the end of 2025, but Suno did not notify its users and still refers to it as a "limited security incident that was quickly contained." The materials only surfaced publicly now, in July 2026 — and the main value of the leak lies not in the personal data, but in the comments within the code, which list the sources and volumes of the training dataset.
Figures They Didn't Want to Show
Previously, Suno had used vague formulations — stating that training was conducted on "essentially all music files of acceptable quality available on the open internet," which amounted to "tens of millions of recordings." The leaked code replaced this vagueness with specifics. Here’s what was actually fed into the model:
- YouTube Music — 2,013,545 music videos, 113,879 hours; a separate labeled corpus "YTM Tagged" — an additional 152,162 hours;
- Pond5 (stock music) — 62,117 hours;
- IMSLP (International Music Score Library Project) — 19,514 hours;
- Genius (song lyrics) — 17,615 hours;
- Deezer — 12,287 hours;
- Jamendo — 3,726 hours;
- Freesound — 410 hours;
- MuseScore — lyrics and scores;
- podcasts via RSS — the code identified around 420,000 podcasts with five or more episodes longer than 30 minutes, aiming to extract about 1 million hours of audio.
The comments in the code leave no room for interpretation: the system pulled data from "genius_hq, youtube_music, freesound, jamendo, imslp, deezer, ytm_tagged," while "non-musical content was filtered out" during post-processing. These are not gray areas or random uploads — this is a designed data collection pipeline from all available web sources.
Why This Is Impossible Without Proxies
A key detail that is easy to overlook amid the impressive figures: YouTube Music, Deezer, and most streaming platforms aggressively defend against mass downloading. Strict request limits per IP, behavioral and TLS fingerprinting, CAPTCHAs, range bans — all of which we regularly discuss. Downloading two million videos and hundreds of thousands of hours of audio from one or even a hundred addresses is impossible: the platform's infrastructure will block the source after the first few thousand requests.
To keep a pipeline of this scale running for months, a rotation of thousands of different IPs that appear as real users is necessary — meaning residential and mobile proxies. Data center addresses on popular streaming services are filtered out almost immediately by ASN. The downloading mechanism itself is typical — a combination like yt-dlp with a proxy layer; we detailed this in our guide on how to download from YouTube using yt-dlp and proxies to bypass blocks. The Suno leak is a rare case where the scale of this "invisible foundational layer" is visible from the outside: the model itself is a showcase, while beneath it, a proxy farm has been running for years, extracting content on an industrial scale.
This also explains why the proxy market in 2026 began to be viewed as infrastructure for AI, rather than a niche tool for scrapers. Any large training dataset — whether musical, textual, or video — is primarily about the logistics of collection: thousands of IPs, evading detection, load distribution. If you want to understand this layer more deeply, we have a separate analysis on proxies for AI training and dataset collection.
Legal Minefield: It's Not About Proxies, But What and Where
Suno has previously been sued by major labels, and the leak strengthens their position. The company defends itself with the fair use doctrine: claiming that training on publicly available music is legal. However, copyright holders have a more serious trump card — DMCA. Their argument: scraping from YouTube on an industrial scale means bypassing technical protection measures that the platform has put in place against automated downloading. Bypassing such measures constitutes a separate violation, regardless of how the fair use dispute ultimately resolves. Similar lawsuits are being pursued against Suno's competitor — the Udio service, which is accused of similar scraping from YouTube.
The takeaway for anyone collecting data is the same as in the case of the EDPB guidelines on scraping under GDPR: proxies solve the technical access problem, but do not provide a legal basis. A residential IP makes you look like a regular visitor — it does not negate the terms of service, copyright, and DMCA. The legal perimeter is defined by WHAT and WHY you are collecting, not by how clean your traffic appears.
What to Take Away from This
Even if you have never collected music for AI, the Suno case offers a concentrated lesson on data collection in 2026:
- Scale relies on proxies. Any serious collection from streaming and protected platforms requires residential or mobile IPs with rotation — data centers are filtered out by ASN within minutes. The type of proxy is chosen based on the goal: residential for the broader web, mobile for the most aggressive anti-bot systems.
- Technical access ≠ rights. Bypassing anti-scraping protections can be classified as a DMCA violation and a breach of terms of service. Separate open data without technical barriers from protected content — these are different levels of risk.
- Proxy hygiene and origin matter. In light of the shutdown of botnets like NetNut, the origin of residential IPs has become a selection criterion: a clean pool from a transparent provider offers both stability and reduced legal risks.
- Data leaks along with methods. Suno denied direct scraping for years — and one hack nullified all those statements. Build your pipeline as if it will be scrutinized in court tomorrow.
Conclusion
The Suno hack is interesting not for the personal data, but because it revealed for the first time the inner workings of a large AI product in specific numbers: 2,013,545 videos from YouTube Music, dozens of sources, a million hours of podcasts planned. Behind these figures lies not the magic of neural networks, but the mundane logistics of collection: a proxy farm, evading detection, months of extraction. The AI race of recent years is largely a race for data, and the race for data is a race for access infrastructure. Suno just happened to be the one from whom this infrastructure was observed.
If you are building a legal data collection operation — for analytics, model training, or monitoring — start with the right foundational layer: clean residential and mobile proxies with transparent origins. The access technique is solvable; the key is to keep it within the bounds of the law and common sense.