Meta has put new limits on the AI coding tools its own engineers can use, aiming to keep output from competitors out of the data it uses to train its models.
According to The Decoder, Meta is restricting its engineers' use of Anthropic's Claude and OpenAI's Codex specifically to prevent the output of those tools from being incorporated into Meta's own training data.
The restriction is targeted rather than blanket. According to a report surfaced via Bing News on MSN, engineers in Meta's Applied AI team can no longer use Anthropic's Claude Code or OpenAI's Codex without approval, a move tied to what the report describes as distillation risks.
"Distillation" refers to using one AI model's responses to help train another. If a company's engineers lean on a rival's tool while building or refining their own systems, that rival's output can seep into the training pipeline. That raises both technical concerns about model quality and legal or contractual concerns, since major AI providers generally prohibit using their outputs to build competing models.
The practical effect is that Meta staff who want to use these particular outside coding assistants now need sign-off rather than free rein. The Decoder frames the policy as a guardrail meant to keep a clean separation between what rival tools produce and what Meta feeds into its own model development.
The sources here are brief, and they do not specify how many engineers are affected, when the policy took effect, or whether it extends beyond the Applied AI team.
Why it matters: the move is a small but telling sign of how guarded the leading AI labs have become about their training data, treating even everyday developer tools as potential leaks in a fiercely competitive race.