Meta Restricts Engineers’ Use of Claude Code And Codex Over Model ‘Distillation’ Concerns

Meta Platforms has instructed engineers in its Applied AI division to limit or restrict their use of Anthropic’s Claude Code and OpenAI’s Codex coding and agent tools, according to internal documents reviewed by The Information. The policy, driven by concerns over inadvertent model distillation, aims to prevent outputs from rival AI systems from contaminating Meta’s own training data and model development processes for its Llama family of models (which, quite frankly, could only help).

The move reflects the increasingly zero-sum nature of frontier AI development, where companies aggressively protect the provenance and purity of their training data while seeking to reduce reliance on competitor tools. Internal guidelines referencing the restrictions date back to at least May, with the policy actively in effect as of late June. Meta has not publicly confirmed or commented on the directive.

According to the internal documents, strict limits have been placed on how engineers in the applied AI division can use the rival tools. The stated goal is to block “inadvertent distillation” of competitor model outputs into Meta’s AI development pipeline. The scope is targeted: it focuses on engineers working directly on model building and applied AI initiatives rather than the entire engineering organization.

Claude Code from Anthropic and Codex from OpenAI are basically the industry standard now for professional developers engaged in agentic coding workflows. These desktop and app-based interfaces can plan, write, debug, and iterate on complex codebases, offering powerful assistance at relatively low individual subscription costs. That accessibility, however, has increased the potential surface area for the risks Meta is now seeking to contain.

What “Distillation” Means

Model distillation is a well-established technique in which outputs from a larger or more capable “teacher” model are used to train or improve a “student” model. In this instance, Meta is concerned that high-quality code suggestions, architectural recommendations, debugging logic, and reasoning traces generated by Claude or Codex could be incorporated – whether intentionally for productivity or accidentally through copied artifacts – into internal codebases, documentation, or synthetic training data.

The result would be a subtle transfer of competitor capabilities into Llama models. Beyond intellectual property exposure, the risk includes contamination of Meta’s carefully curated training data pipelines and the creation of unintended dependencies on rival model behaviors. Secondary concerns involve proprietary Meta code and context being transmitted to external Anthropic and OpenAI servers during routine usage.

The move comes as Meta is locked in a high-stakes competition to close the capability gap with OpenAI, Anthropic, and Google – while simultaneously constructing massive internal infrastructure. The company has publicly emphasized its desire to reduce dependence on third-party AI services for both cost and strategic autonomy reasons. Restricting these widely used coding tools sends a clear internal message: engineers should build with Meta tools and data wherever possible.