Tarun Chhabra warns distillation is a national security risk, but Silicon Valley keeps importing

When Anthropic’s chief national security officer and former Biden administration export control architect Tarun Chhabra highlighted model distillation as an emerging national security concern, the point was not subtle. In the national security framing, “distillation” is not just a technical optimization. It is a pathway that can help AI capability move faster, and with fewer visible traces, than traditional “build it from scratch” approaches.

For U.S. policymakers trying to contain China’s AI lead, that is the hard part. Even as lawmakers and regulators try to tighten export controls around the most sensitive pieces of the AI supply chain, the real-world ecosystem is full of developers and companies adopting Chinese AI models and techniques. In other words, the U.S. can put friction in one lane, but innovation can still flow in adjacent lanes. Chhabra’s warning puts a spotlight on one of those lanes: distillation, which can compress knowledge from larger models into smaller ones.

This is where the story stops being a policy debate and starts looking like an operational problem for everyone who touches AI. The report frames a broader dynamic: while the U.S. wants to protect its lead, Silicon Valley keeps using Chinese AI models. It also notes that Apple, Thinking Machines, and developers worldwide have been embracing Chinese AI models such as Kimi K3. That matters because these are not fringe experiments. When mainstream product teams and well-known AI builders incorporate models developed elsewhere, containment strategies lose some of their ability to shape actual adoption.

Model distillation itself is at the center of the tension. Chhabra flagged it as an emerging national security concern, and the underlying logic is straightforward. Distillation can produce a smaller model that mimics the behavior of a larger one, which means capabilities can be repackaged. If policymakers are trying to slow down the spread of frontier capabilities through export controls, a technique that changes how capability is transferred, and how it is counted or detected, becomes strategically relevant. It is less about who holds the largest model and more about who can deploy useful intelligence quickly.

Now add the policymaker reality: export control regimes tend to lag behind how products are built. The report emphasizes that U.S. policymakers are struggling to protect the AI lead, even as more companies and developers look to Chinese models. That mismatch is what makes the issue feel especially brittle. Controls can be written to target specific inputs, specific workflows, or specific compute pathways. But the AI world keeps redesigning workflows. Distillation is one example of the broader pattern: the technique might not be the thing being controlled, but it can still change the outcome.

The market incentives do not help. Developers want working systems, fast. Enterprises want performance and cost efficiency. Creators want tools that reduce experimentation time. When Chinese AI models such as Kimi K3 are available and getting adopted, it is not automatically because teams ignore U.S. policy. It is because product teams live in a world of deadlines, user demand, and competitive pressure. If you can ship an assistant that feels better because you used a model that already has it, you are unlikely to pause just because containment is a headline on Washington’s calendar.

Chhabra’s credibility matters here, because the report ties him to the government side of the equation. He is not only an AI executive, he is Anthropic’s chief national security officer, and he was previously an export control architect in the Biden administration. That background is why his callout of model distillation lands. It signals that the concern is not speculative chatter. It is a way of thinking about control strategies that are no longer just about chips and training runs, but also about the downstream techniques that redistribute capability.

For decision-makers, the second-order implication is uncomfortable: compliance and strategy might diverge. Boards and executives can invest in governance, but they still have to decide what to do when the easiest route to product performance relies on technologies that policymakers are trying to constrain. The report frames this as an ongoing struggle, and that struggle likely does not get easier as techniques like distillation spread and as adoption of Chinese models continues.

If you are an AI operator, a product leader, or a funder with exposure to infrastructure and models, this is the moment to internalize the lesson: containment is not one wall, it is a maze. The U.S. can reroute traffic, but it cannot assume teams will stop moving. Chhabra’s warning about distillation highlights a likely future of enforcement pressure, technical scrutiny, and more complicated compliance math, particularly for anyone building systems where model choice and model transformation workflows determine the final capability delivered.