Anthropic is done playing nice. The AI safety company is demanding legal action against Alibaba, accusing the Chinese e-commerce giant of orchestrating the largest known cloning attack on its flagship model, Claude. According to a bombshell report filed by Anthropic, Alibaba allegedly used a staggering 25,000 accounts to extract 28.8 million exchanges from Claude over an extended period. That’s not research. That’s industrial-scale theft.
Inside the Heist
The attack was brazen. Alibaba’s accounts systematically queried Claude across millions of sessions, harvesting responses to train a rival model. Think of it as a digital bank robbery where the thieves didn’t just grab cash—they copied the vault’s blueprint. Anthropic’s internal logs show the extraction happened in waves, with accounts rotating to avoid rate limits. The company says the data haul is equivalent to years of human-generated conversations, enough to replicate Claude’s conversational DNA.
“This isn’t a few bad actors experimenting,” an Anthropic spokesperson told me. “This is a coordinated corporate campaign to steal our intellectual property. Alibaba must be punished.”
Why This Matters More Than You Think
AI models are expensive. Claude costs millions to train, and its safety alignment—the guardrails that keep it from spewing hate or giving bomb recipes—takes even more. When a rival clones your model, they don’t just steal your technology; they steal your trust. Alibaba could now have a black-box copy of Claude, stripped of the safety constraints that Anthropic spent years refining. That’s a nightmare scenario for AI governance.
Anthropic has been the darling of the safety crowd. Founded by ex-OpenAI defectors, it positioned itself as the responsible alternative—the company that would build AI without burning the world down. But Alibaba’s attack reveals a vulnerability: the very openness that makes AI accessible also makes it stealable. Every API call is a potential data point for a competitor. The industry’s business model—selling access to models—is built on trust. Alibaba just showed how fragile that trust is.
The Scale Is Unprecedented
Twenty-five thousand accounts. Twenty-eight point eight million exchanges. Let’s put that in perspective. The entire Common Crawl dataset—a massive web scrape used to train models—contains about 200 billion tokens. Alibaba’s haul from Claude is a significant fraction of that. It’s enough to fine-tune a competitive model or bootstrap a new one from scratch. The attack likely cost Alibaba a few thousand dollars in API fees. The value of what they took? Tens of millions.
This isn’t the first time Chinese firms have been accused of ripping off Western AI. But it’s the largest, and Anthropic is determined to make an example of Alibaba. The company has already notified US regulators and is exploring legal avenues. “We will pursue this to the fullest extent of the law,” the spokesperson said. “No one is above the rules.”
“This isn’t a few bad actors experimenting. This is a coordinated corporate campaign to steal our intellectual property.” — Anthropic spokesperson
The Regulatory Vacuum
The attack exposes a glaring gap in AI governance. There are no clear rules against model cloning. Copyright law is a mess when it comes to training data, and terms of service are often unenforceable across borders. Alibaba likely violated Anthropic’s API terms, but what’s the penalty? A slap on the wrist? A lawsuit that takes years? By the time courts act, the stolen model is already deployed.
Anthropic is pushing for something stronger: a legal precedent that treats large-scale model extraction as theft, not just a policy violation. They want courts to recognize that an AI model’s behavior—its unique outputs—is protectable intellectual property. If they win, it could reshape how AI companies defend their models. If they lose, expect a feeding frenzy.
Alibaba’s Silence
Alibaba hasn’t responded to the allegations. Not a peep. That’s telling. If the accusations were false, you’d expect a swift denial. Instead, radio silence. The company’s AI division, Damo Academy, has been quiet since the report dropped. Insiders say the legal team is scrambling. This story is far from over.
The Bigger Picture
Anthropic’s fight with Alibaba is a microcosm of the AI arms race. The US and China are locked in a battle for dominance, and intellectual property is the new oil. But there’s a twist: the more we lock down AI, the less we can trust it. Safety research depends on transparency. If every company builds a walled garden, we’ll never know if models are safe. Alibaba’s theft is illegal, but it also highlights how porous the current system is.
Anthropic has a choice to make. It can fight this battle in court, setting a precedent that may chill legitimate research. Or it can accept the loss and advocate for stronger industry norms. The company seems to have chosen the first path. Good. Someone needs to draw a line.
You want to know what happens next? Watch the courts. Watch China. And watch your back if you’re an AI company. Because if Alibaba gets away with this, every model is up for grabs.
