Meta Restricts Engineers' Access to Claude Code and Codex Over Model Distillation Fears

Meta has ordered engineers working on the development of its own AI models to limit their use of Anthropic's Claude Code and OpenAI's Codex. The company fears that outputs from rival tools could unintentionally end up in the training data for its own systems, risking a phenomenon known as model distillation.

Model distillation involves one system learning from responses generated by another, stronger model. A smaller or weaker model thereby absorbs the capabilities of the larger one, without needing to be trained from scratch on the original data. It's a technique widely used across the industry, but controversial when it happens without the consent of the original model's owner - exactly what Anthropic recently accused entities linked to China's Alibaba Qwen of doing.

What exactly was banned

According to an internal memo, some teams at Meta were instructed to halt tasks that rely on external tools. Employees can still use rival models for routine tasks, but are barred from using AI-generated content to create training or evaluation data for Meta's own models.

Serious escalations with partner companies if their models' outputs leaked into Meta's training data - from an internal Meta memo

The scope of the restriction is deliberately narrow. It mainly applies to engineers working directly on model development and Applied AI initiatives, not Meta's entire engineering organization, where tens of thousands of developers work day to day.

The race for an in-house tool

The restrictions coincide with the development of MetaCode, an internal coding tool meant to eventually replace Claude Code and Codex in the daily work of Meta's teams. The company has spent months investing in its own coding models as part of a broader strategy to reduce reliance on external AI vendors, a push that also fits into Meta's previously announced big leap in coding and agentic work built around the Muse Spark model.

Meta's decision shows that concerns about model distillation are no longer purely a geopolitical issue between the US and China. Increasingly, they also touch on rivalry among American tech giants, which simultaneously use each other's tools while building their own competing versions, trying to prevent knowledge from leaking in either direction.

What it means for companies and developers

For companies using coding agents in their daily work, Meta's case is a signal that even the largest organizations treat outputs from external AI models as a potential risk to their own intellectual property. Polish tech companies increasingly rolling out Claude Code, Copilot, or Codex within their development teams might consider a similar separation between operational use and using outputs to train their own models or build internal knowledge bases.

The case also touches on a broader question of who actually controls the data generated while working with commercial AI agents, and where the line lies between using a tool and inadvertently giving competitors insight into one's own processes.

What's next

Meta has not officially commented on the details of the internal guidelines. The pace of MetaCode's development, and whether the company decides to extend the restrictions to a wider group of engineers, will show how seriously it takes the risk of distillation compared with the benefits of using mature, external coding tools.

Sources: Meta restricts use of Claude Code and Codex to keep rival AI out of its training data (the-decoder.com), Meta restricts engineers' use of Claude Code and Codex over model 'distillation' concerns (zerohedge.com), Meta Platforms restricts engineers from using Anthropic's Claude and OpenAI's Codex (cryptobriefing.com)