Remember when "model extraction" sounded like a sci-fi heist — sneaking into a server room, downloading the brain of an AI in a single dramatic USB-stick moment? Forget it. The real version is far less cinematic and somehow more unsettling: millions of polite, perfectly ordinary-looking questions, fired from thousands of accounts, slowly reverse-engineering everything that makes a model tick. No trench coats required.

That's the scene Anthropic says it walked into on 25th June 2026, when it publicly accused operators linked to Alibaba of running what it's calling the largest known attempt yet to extract the capabilities of its Claude models. Not a hack or a leak. Just an enormous, patient pattern of interactions designed to learn how Claude thinks well enough to rebuild it elsewhere.

Here's the actual substance behind the story. Anthropic alleges that accounts connected to Alibaba conducted a systematic, large-scale effort to probe Claude's models, using huge volumes of queries across thousands of separate accounts to map out its capabilities and, effectively, copy its know-how without copying its code. Anthropic is framing this as the biggest case of its kind so far, a milestone nobody wanted but a milestone nonetheless.

What makes this notable isn't just the scale (though "millions of interactions" is a lot of homework to mark). It's what it signals about where the AI race has quietly moved. For the last few years, the competition was simple, if expensive: build the smartest model. Now there's a second front opening up — defending the smart model you already built from being siphoned out through the front door, one innocuous-looking chat at a time.

If you're running a business that leans on AI tools (and at this point, who isn't), this matters more than it might first appear. The competitive moat in AI was always assumed to be the model itself — the training, the compute, the secret sauce. This case suggests that moat has a leak, and patching it is now its own arms race, separate from the race to build the next great model.

For founders and execs, the takeaway isn't "go audit your AI vendor's security policy this afternoon" (please don't email them in a panic). It's broader than that: the value you're building on top of AI tools — your prompts, your workflows, your proprietary use cases — sits in an ecosystem where the underlying intelligence itself is now considered worth extracting at industrial scale. That should change how you think about what's defensible in your own AI-powered processes, because if the foundation models are being targeted this aggressively, the assumption that "the tech layer is someone else's problem" gets shakier by the day.

It's also a tidy reminder that the AI industry's competitive dynamics are starting to look less like a tech race and more like, well, an actual geopolitical one with accusations, denials, reputational stakes, the works. Building a great product was never going to be the hard part forever; protecting it was always going to catch up as the next battleground.

So there you have it: the AI wars have officially entered their "industrial espionage" arc, and somehow that was always going to happen eventually (every gold rush gets its claim-jumpers). Anthropic built something good enough that someone allegedly tried to copy it at scale rather than build their own — which is, in its own backhanded way, the loudest compliment in tech right now.

Just maybe don't put that on the slide deck.

— The Business Index Team