Meta Is Scared Its Own Engineers Will Leak Claude's Brain Into Llama
Meta quietly restricted its AI engineers from using Claude Code and Codex over distillation fears. The AI industry's open secret is becoming its biggest legal battlefield.
Meta Is Scared Its Own Engineers Will Leak Claude's Brain Into Llama
Here's something you don't see every day: one of the largest AI companies on earth is restricting its own engineers from using rival AI coding tools — because it's afraid those tools might accidentally make Meta's models too smart.
Meta, according to internal documents obtained by The Information, has placed strict limits on how engineers in its applied AI division can use Anthropic's Claude Code and OpenAI's Codex. The company has even temporarily halted certain work with these models. The reason? A single word that's becoming the most loaded term in AI: distillation.
What Meta Is Actually Doing
Let's be specific. Meta isn't banning Claude Code or Codex company-wide. Engineers can still use them. But the applied AI division — the team building Meta's own models — now operates under rules that would make a compliance officer nod approvingly:
- Engineers cannot use AI outputs to create test tasks or training data
- Code analysis generated by rival models must not flow into Meta's training pipelines
- Human review is required before any AI-generated code touches production
- Certain work with Claude Code and Codex has been temporarily paused entirely
An internal memo warned of "serious escalations with partner companies" if rival model outputs leaked into Meta's training data. That's corporate-speak for "Anthropic and OpenAI will sue us into oblivion."
Meta is simultaneously building its own coding assistant, called MetaCode, to cut reliance on outside tools. Partly because of distillation risk. Partly because of cost — the company is on track to spend billions on internal AI usage this year alone.
What Is Distillation, Really?
If you've been around AI for a while, you know this term. But it's worth explaining because it's the crux of why this story matters.
Distillation is when you use a powerful model's outputs to train a smaller, cheaper model. You send prompts to GPT-5 or Claude, collect the responses, and use those input-output pairs as training data for your own model. The student model learns to mimic the teacher's behavior without ever seeing the teacher's weights.
It's legal in some contexts. Apple has an arrangement with Google to distill Gemini into on-device models. Amazon has certain rights to distill Anthropic's models. These are negotiated, contractual relationships.
But unauthorized distillation? That's the AI industry's equivalent of insider trading — everyone suspects everyone else is doing it, nobody wants to get caught, and the penalties are potentially enormous.
OpenAI, Anthropic, and Google all explicitly ban using their model outputs to build competing systems in their terms of service. The logic is simple: if you could just query Claude a billion times and train Llama on the responses, you'd be getting Claude's capabilities for the price of API calls. That's not a business model any frontier lab can survive.
Everyone Is Doing It (Or Accused of Doing It)
The Meta story doesn't exist in a vacuum. Distillation is the open secret of the AI industry right now.
In April, Elon Musk had to admit that xAI partially distilled OpenAI's models. Not "was accused of" — admitted. The man who founded OpenAI and then sued it was caught using its outputs to train Grok. You can't make this up.
Anthropic recently accused Alibaba of "the largest known distillation attack to date." Whether that's hyperbole or fact, it signals that frontier labs are now actively hunting for evidence their models are being siphoned.
And Amazon — Anthropic's own investor, having poured up to $25 billion into the company — is reportedly distilling Anthropic's models to cut costs before new token-based pricing kicks in next year. Even your investors want to eat your lunch.
The pattern is clear: the gap between frontier models and everyone else is narrow enough that distillation is worth the risk, and wide enough that frontier labs will fight to protect it.
Why Meta Specifically Is Paranoid
Meta's situation is uniquely awkward. They're the only major AI lab that open-weights its frontier models. Llama weights go out to the world for free. That means Meta can't exactly complain if someone distills Llama — they gave it away.
But when it comes to Claude and Codex, Meta is on the other side. It's a consumer of closed models, not a provider. And its engineers are using those closed models in the same codebase where Llama is being developed. The risk isn't theoretical — if an engineer uses Claude Code to generate a function, and that function ends up in a training dataset, Anthropic has a legitimate claim that Meta stole its IP.
Meta's paranoia is actually rational. The company that open-weights everything has the most to lose from a distillation scandal because it would undermine their entire "open AI" narrative. Imagine the headlines: "Meta, Champion of Open Source AI, Caught Stealing From Closed Competitors." That's a PR disaster that makes the Cambridge Analytica coverage look like a rounding error.
The Real Question: Can You Even Prevent It?
Here's what I find most interesting about this story. Meta is trying to prevent distillation through policy — memos, rules, human review requirements. But can you actually stop it?
Think about the workflow. An engineer opens Claude Code. Claude generates a solution. The engineer reads it, understands it, and writes their own version. Is that distillation? What if they copy-paste? What if they refactor? What if the AI's approach influences their thinking but the final code is original?
This is the AI equivalent of the clean room design problem in intellectual property law. You can't unsee what you've seen. Once an engineer has been exposed to Claude's approach to a problem, their subsequent work is — in some hard-to-quantify way — influenced by it.
Meta's policy of requiring human review is a good-faith effort, but it's fundamentally a speed bump, not a wall. The only real defense is what Meta is already doing: building MetaCode so engineers never need to leave the Meta ecosystem. If your own tool is good enough, the distillation risk disappears because there's nothing to distill from.
But that's a big "if." MetaCode isn't there yet. And until it is, Meta's engineers are caught between using the best available tools (Claude Code, Codex) and following rules that say "don't let the good stuff leak into our stuff."
The Bigger Picture: AI's Original Sin
Distillation is AI's original sin. The entire field was built on training on data that belonged to other people — books, articles, code, images. The models got good by absorbing the internet's collective output. Now that the models are valuable, their creators want to prevent the same thing from happening to them.
There's a delicious irony here. OpenAI trained on GitHub repos without asking. Anthropic trained on web content without asking. Meta trained on everything without asking. Now they're all terrified that someone will train on their outputs without asking.
The difference, of course, is that training on the internet produced something new — a model with emergent capabilities. Distillation produces a copy. It's the difference between reading a textbook and photocopying the exam answer key. One is learning. The other is cheating.
But the line between them is blurrier than the frontier labs would like to admit. If an engineer at Meta learns a better coding pattern from Claude Code and applies it in Llama's training data, is that learning or cheating? The law doesn't have a clean answer. Neither does Meta's compliance team, apparently.
What This Means For You
If you're building AI products, this story is a preview of the regulatory and legal landscape that's coming. Distillation clauses in terms of service are going to become as important as GDPR clauses in privacy policies. If you're using Claude or GPT APIs to generate training data — even indirectly — you need to think about whether your usage agreement allows it.
If you're using open-source models (like we do at CopperRiver with GLM, DeepSeek, and Qwen), you're in a safer position. Open-weight models don't have distillation restrictions because they're already open. You can fine-tune them, distill them, modify them. That's the point.
And if you're an engineer at a big AI lab — maybe read the memo before you fire up Claude Code tomorrow.
The Takeaway
Meta restricting its own engineers from using rival AI tools is a symptom of a deeper problem: AI capabilities are now valuable enough to steal, and the industry hasn't figured out how to protect them. Policy memos and human review requirements are band-aids. The real solution — building tools good enough that you don't need anyone else's — is still a work in progress.
The distillation wars are just getting started. Meta's internal restrictions are the opening move. Expect more companies to follow, expect more accusations to fly, and expect the line between "learning from" and "stealing from" to get a lot more attention in courtrooms.
For now, the most open AI company in the world is quietly telling its engineers: don't let the closed models rub off on you. Make of that what you will.
Want to use AI models that don't come with distillation paranoia? CopperRiver runs on open-source models like GLM, DeepSeek, and Qwen — no API lock-in, no terms-of-service landmines. Just a desktop AI assistant that actually gets things done.