Meta restricted engineers' use of Claude Code and Codex due to distillation risk

Meta has introduced strict restrictions on engineers in its applied AI subdivision using code-writing tools Claude Code from Anthropic and Codex from OpenAI — according to reporting by The Information citing sources.

Why Meta needs its own code tools

The company wants to develop its own tools for automating code writing and simultaneously reduce its engineers' dependence on competing products they use today. While Meta works on built-in alternatives to Claude Code and Codex, management is asking engineers in the applied AI subdivision to be more careful with external tools.

Claude Code is an agent from Anthropic that autonomously works with the codebase through the terminal, relying on models from the Claude family. Codex is OpenAI's similar product, integrated with GPT models and solving similar code automation tasks. Both tools have over the past years become some of the most notable AI agents for writing and refactoring code on the market.

What is the risk of distillation

The officially stated reason for Meta's restrictions is the risk of inadvertent distillation. This is what the industry calls a situation where active use of a third-party AI tool can inadvertently pass signals to a competitor about the company's internal processes, code patterns, or approaches — and thereby help Anthropic or OpenAI improve their own models at Meta's expense.

• Restrictions affected engineers in the applied AI subdivision at Meta
• The restrictions targeted two of the most prominent AI agents for code on the market — Claude Code (Anthropic) and Codex (OpenAI)
• The official reason for the restrictions is the risk of inadvertent distillation
• The information was first published by The Information citing sources
• In parallel, Meta is developing its own code-writing tools to reduce dependence on competitors' products

Who else is involved in this race

The situation is instructive in that the largest AI labs simultaneously serve as suppliers and competitors to each other. Anthropic and OpenAI sell access to their development tools to companies like Meta, which in parallel competes with them on the market for base models and intends to offer the market its own AI products for programming. Such mutual dependence is typical for the entire industry: almost every major technology company simultaneously uses other companies' AI services internally and develops competing solutions, making questions of protecting internal data and developments especially sensitive precisely in the internal divisions where a company's most valuable products are created.

Why this is sensitive right now

The race for the best AI tool for developers is happening in parallel with the race for the best large language model — and this is no coincidence. Engineers working on the most complex and closed parts of the codebase daily formulate detailed technical tasks for external AI agents, show them fragments of internal architecture, and receive generated code in return. For a company of Meta's scale, which simultaneously invests billions in its own AI models and remains a customer of its direct competitors' products, this daily routine becomes a matter of information security no less important than protecting source code or user data.

What this means

Meta's restrictions show that large technology companies are starting to treat third-party AI tools for development not just as a means of improving engineer productivity, but also as a potential channel for leaking competitive advantages. As AI agents for code become an integral part of daily developer work, the question of who ultimately gets the data about how code is written inside a company becomes increasingly strategically important — especially for a player of Meta's scale, which simultaneously serves as both a customer and a direct competitor of Anthropic and OpenAI on the AI market.