The code does not lie; only the founders do. Sony Music and Warner Music just pulled off the cleanest exploit of the year, and Anthropic is the target. The two record labels filed a copyright infringement lawsuit against the "safety-first" AI company, claiming Claude, its large language model, reproduces copyrighted song lyrics without authorisation.
This is not a legal dispute. This is an attack vector. And Anthropic walked into it with the same careless confidence of a DeFi protocol that ships unaudited code.
I have spent the last five years auditing smart contracts for flash loan attacks, reentrancy bugs, and governance exploits. This lawsuit is a familiar pattern: a project that prioritises growth over security, that treats external data as free labour, and that assumes goodwill will substitute for access control. In the blockchain world, we call that a rug pull waiting to happen.
Now, let me dissect the mechanics.
Context
Anthropic is one of the most richly funded AI companies in the world, with a valuation reportedly north of $60 billion in its latest rounds. Its core product, Claude, is an enterprise-grade AI assistant. The company's entire brand is built on responsible AI. It talks about alignment, interpretability, and avoiding harmful outputs. Yet its models memorise popular song lyrics, and its data pipeline has no licensing layer.
Sony and Warner Music, two of the three major record labels, decided to test that brand. The lawsuit alleges that Claude generates verbatim or near-verbatim lyrics from songs in their catalogues, ranging from Katy Perry to Beyonce. This is not a hallucination. The model can be prompted to recall lyrics because it was trained on them. The technical term is "memorization," and it is the dirty secret of the LLM industry.
Let's be clear about the technical reality. Anthropic's training data comes from a mix of public web scrapes, including Common Crawl, which contains thousands of lyrics websites such as Genius and AZLyrics. Those sites hold the text of essentially every popular commercial song ever released. If you train on that data, your model will memorise the most repeated sequences. Popular song lyrics are among the highest-frequency text patterns on the internet. No amount of reinforcement learning from human feedback will erase them.
Anthropic's own model card does not disclose a quantitative assessment of lyric memorization or any mitigation against it. That is not an omission; it is a signal. They didn't know. Or they didn't want to know. In my security audits, I see the same failure: a protocol that skips the reentrancy check because the attacker would be a "good actor."
Core Analysis
This lawsuit is best understood as a reentrancy attack on the training data's state. In a smart contract, reentrancy occurs when an external call is made before the contract's internal state is updated. An attacker can call back into the function repeatedly and drain funds. Here, the "funds" are the intellectual property encoded into the model's parameters. The model is the smart contract. The training data is the external call. Anthropic never updated the "licensing state" before ingesting the data. So the labels called back, and the model emitted the lyrics. Simple. Predictable. Expensive.
The technical evidence is straightforward. The model's output is deterministic for high-memorization patterns. If you prompt Claude with "Can you finish these lyrics?" for a popular song, it will likely comply. This is the AI equivalent of an unpatented vulnerability: publicly exploitable, widely known, and not fixed. Sony and Warner likely used adversarial extraction techniques, such as repeated prefix attacks, to collect a clean evidence chain. Their lawyers now hold a concentrate of model outputs that match the original lyrics with high similarity.
Now, here's where the crypto mindset matters. In the blockchain space, we talk about provenance. Every token, every NFT, every transaction can be traced to its source. AI training data has no such traceability. You cannot point to a specific song and say, "That exact line came from that exact file." The data is mashed together, shuffled, and transformed into weights. This is why Anthropic's first defence will be "we can't prove the model copied, it just learned patterns." That defence works in law school, not in engineering. The model learned the pattern because the pattern was repeated in the data. That is proof.
I don't trust audits; I trust the gas fees. In the audit world, we know that a protocol can pass a hundred checks and still get hacked. The only thing that matters is the incentive structure. Anthropic's incentive was to build the most capable model as fast as possible. Licensing music would have slowed them down and reduced their training data diversity. So they skipped it. The result is a latent liability buried in billions of weights.
But the damage goes beyond Anthropic. This lawsuit is the first coordinated attack by a concentrated, data-rich industry against the AI training supply chain. The music industry has something that the news industry and the book publishing industry lack: an organised, centralised copyright database. Performance rights organisations like ASCAP and BMI have been collecting royalties for decades. They know exactly who owns what, and they have the legal infrastructure to act as a cartel. The labels are not simply defending creators; they are defending their position as toll-booth operators on cultural production. This is a systemic incentive dissection: the plaintiffs are not only claiming past harm, they are demanding a percentile cut of AI revenue going forward.
If they win, the cost of training data skyrockets. Small AI startups that rely on open scrapes become instantly illegal. This is a classic winner-takes-all move: large players can absorb licensing fees; businesses with less capital cannot. The same thing happened in DeFi: when regulation hit, small protocols died, and the big ones hired compliance teams.
Contrarian Angle
Now let me steelman the bulls' case. The fair use defence is not laughable. The US Copyright Act allows "transformative" uses. Anthropic will argue that training a neural network is transformative, that the model does not replace the original song, and that the public benefits from AI's ability to analyse and quote lyrics. Courts have historically allowed transformative uses for search engines and Google Books. And the music industry is no saint. The labels have a long history of underpaying artists while profiting from their work. In this lawsuit, the actual songwriters may not see a dime; the corporate parents will fight for their share of the AI pie.
But here is the counter-intuitive blind spot: the labels are also exposing themselves. If the court rules that training requires explicit permission, then every piece of public data comes under suspicion. Not just lyrics, but every article, every comment, every code snippet. That would crush the open-source AI movement. Meta's Llama and Mistral would face class actions. The entire open internet becomes a minefield.
So the bulls are right: this could be a stability risk for the industry. But that risk is exactly why I shorted the narrative. Markets hate uncertainty. Until a judge issues a ruling, every AI company is walking around with a reentrancy bug in its training data. And any plaintiff can trigger it.
Takeaway
The rug was pulled before the mint even finished. Anthropic's compliance failure is not a bug; it is a feature of a trust-based system that assumed creators would never fight back. The lesson for crypto and AI alike is that you cannot outsource governance to a whitepaper. You need on-chain provenance, cryptographic hashes of licensed data, and smart contracts that automatically enforce royalty payments. Without that, your model is not an innovation; it is a liability waiting to be called. And the calling party is already at the door.
