OpenAI, Anthropic, and Google unite to fight model distillation

Silicon Valley's biggest rivals are officially joining forces to stop what they see as systemic intellectual property theft.

OpenAI, Anthropic, and Google have started actively collaborating to detect and curb unauthorized model copying by Chinese competitors, Bloomberg reported.

To coordinate, the three companies are sharing operational threat intelligence through the Frontier Model Forum, the industry nonprofit they co-founded with Microsoft back in 2023.

Borrowing the cybersecurity playbook

The target of this alliance is adversarial distillation, a method where developers query a frontier AI model millions of times and use its outputs as synthetic training data for cheaper clone systems.

Stanford's Alpaca experiment proved years ago that this technique worked, but it has since matured into an industrial-scale operation.

According to Bloomberg, US officials estimate that adversarial distillation costs American AI laboratories billions of dollars in lost revenue every single year.

The alliance is now treating distillation essentially like a cyberattack. Much like enterprise cybersecurity firms routinely swap threat indicators and malware signatures, the three AI labs are exchanging technical data to spot patterns of terms of service violations.

The finger-pointing goes public

This tension has been brewing for months, but the specific targets are now coming into full view.

OpenAI warned the US Congress in February that Deepseek was employing increasingly sophisticated methods to extract data from American frontier models.

Anthropic went further by naming names, identifying Deepseek, Moonshot, and Minimax as entities involved in model distillation practices.

Teaming up and going to lawmakers shows that private account suspensions are no longer enough to stem the tide.

Why it matters

For enterprise customers and developers, tighter defensive measures mean API governance is about to get much more aggressive. Expect stricter verification hurdles, harsher rate limits, and potentially aggressive IP screening that could introduce friction for legitimate power users.

For the broader AI ecosystem, this marks a pivot from a pure model capabilities race to an asymmetric economic war. Top labs spend billions training frontier models; rival teams spend millions querying them to jump years ahead in reasoning benchmarks.

We think borrowing collective threat sharing from cybersecurity is a smart, pragmatic countermeasure that raises the technical friction of mass data harvesting.

However, it looks like a difficult line of defense to hold indefinitely. As long as these models are publicly accessible through commercial APIs, separating a dedicated distillation pipeline from legitimate heavy enterprise usage will remain an agonizingly difficult cat-and-mouse game.