Meta restricts Claude Code Codex engineers distillation training data
Meta has told engineers to limit use of Anthropic's Claude Code and OpenAI's Codex, worried that rival model outputs could seep into its own AI training data through distillation. Internal documents flag legal and partner-contract risk, not just cost. Meta $META is leaning more o
Meta $META has moved to restrict its engineers from using two of the most popular AI coding assistants, Anthropic's Claude Code and OpenAI's Codex, according to internal documents. The concern is not the monthly bill. It is distillation: the risk that outputs from rival models, pulled in while engineers debug or build, could end up embedded in Meta's own training data. When an engineer uses an external coding tool on proprietary AI training infrastructure, chunks of that code travel to outside servers. That is normal for these tools, but it becomes sensitive when the codebase is the company's frontier-model plumbing. Internal language reportedly warned that letting rival outputs seep in could trigger serious escalations with partner companies, which suggests the worry is partly contractual. Meta is also building up its own internal coding tools and wants to cut reliance on outside vendors. The episode is a window into how protective the leading AI labs have become. The same companies that sell coding assistants to everyone else are increasingly wary of using each other's, because at frontier scale the training data is the crown jewel.