Meta limits employee access to Claude Code and Codex amid distillation concerns

Meta has placed strict limits on how engineers in its Applied AI division can use Claude Code and Codex, amid concerns that extensive use of AI coding assistants could inadvertently lead to the distillation of proprietary capabilities from Anthropic's and OpenAI's models.

The growing adoption of AI coding tools has turned model distillation from a theoretical risk into a formal enterprise policy concern at one of the world's largest AI labs by headcount. This marks one of the first documented instances of a major frontier AI lab restricting employee access to rival AI coding assistants on the grounds of preventing model distillation.

What is Model Distillation

Model distillation, also known as knowledge distillation, is an AI training technique in which a smaller, more efficient model learns to replicate the behaviour of a larger, more powerful model. Using a teacher-student framework, the smaller "student" model is trained on the outputs generated by the larger "teacher" model, allowing it to achieve similar performance while requiring significantly fewer computing resources.

More from Tech

The process helps developers create AI models that are faster, cheaper to deploy and capable of running on a wider range of hardware, from servers to smartphones. However, because distilled models inherit many of the capabilities of their larger counterparts, companies developing frontier AI systems have increasingly raised concerns that extensive interactions with proprietary models could enable competitors to recreate key capabilities without direct access to the original models.