Post

Conversation

This is either brilliant or scary: Anthropic accidentally leaked the TS source code of Claude Code (which is closed source). Repos sharing the source are taken down with DMCA. BUT this repo rewrote the code using Python, and so it violates no copyright & cannot be taken down!
Image
David Watson 🥑
Post your reply

You can imagine Anthropic being in a pickle: 1. Do they just leave this, and look the other way, ignoring that it's not exactly fair to transform their code and leave it up there 2. Do they claim copyright applies... but this could be bad for their own business in much bigger
Worth asking btw if code fully written with AI has copyright protection? Anthropic is big in saying most of Claude Code is generated by AI. Not 100%, but close, and the last few months pretty much all. Another interesting copyright question.
Quote
Gergely Orosz
@GergelyOrosz
Replying to @zeeg
Yes, but two things: 1. Does AI-generated code have copyright? Anthropic is big on claiming Claude Code is close to 100% AI-generated, using high-level prompts. 2. Assuming it’s copyrighted (some questions on #1 ofc!) would Anthropic try to prove it, given the obvious backlash
This one is actually clear cut copyright violation. Changing copyrighted material does not void the IP protections and thats all they did. “Could you have implemented this without using the source code” is the easy test
Yes, but two things: 1. Does AI-generated code have copyright? Anthropic is big on claiming Claude Code is close to 100% AI-generated, using high-level prompts. 2. Assuming it’s copyrighted (some questions on #1 ofc!) would Anthropic try to prove it, given the obvious backlash
So the DMCA defense is now just "Your Honor, it's in a different language." International law scholars in ruins.
if copyright applies for derived works based on the source, that could be a coffin in the nail for training LLMs that work just like this... and Anthropic is one of the leading AI labs commercializing LLMs. So it could be tough for Anthropic to argue that copyright applies!
That’s honestly both genius and a little terrifying 😅 Turning leaked logic into a clean Python rewrite to bypass copyright is a wild loophole… Shows how fast things move in AI right now — control is an illusion if the idea is out there.
The DMCA response is the Streisand Effect in a zip file. Anthropic removes the TypeScript. The community rewrites it in Python. The code is now more distributed than before. The only thing the takedown proved is that the source maps should not have been there.
Anthropic just had an all-time self-own. Someone on their team ran a production build of Claude Code, generated a source map (a literal blueprint of their entire codebase), and published it straight to npm for anyone to download. This is the same company that: · Told Congress
Image
Leaked by accident. DMCA'd immediately. Rewritten in Python within hours. This is the fastest open source project in history that nobody at Anthropic intended to create
With novels, you can't use AI to translate someone's book into German, and claim it violates no copyright Why is the same not true for code?
This is wild 😂 Leaked TS code gets DMCA'd instantly, but one dev uses AI to rewrite it all in Python in hours now it's untouchable. 40k+ forks already. Copyright vs AI agents: who wins? Anthropic's probably just gonna ignore it.
Image
Refactoring proprietary TypeScript into Python is a flawless 'Clean-Room' reverse-engineering bypass. Under US Copyright law, Anthropic cannot DMCA the logic or the architecture. However, there is a fatal flaw in this clone. If that Python repository still contains the verbatim
The irony of a coding tool's source getting reviewed by the entire developer community in 24 hours is that it's probably the fastest security audit they'll ever get. Unplanned, but arguably useful.
I wonder if they really needed the source to do that, I bet Claude itself could do a decent job rewriting in Python based on the minified JS. It would be painful and lose a lot of intent but could get pretty far.
Does this mean we can fully automate clean room reverse engineering, but without even needing two separate agents, just give an agent the protected source code, and have it, while being able to directly observe the original, write a re-implementation?
Its a snapshot. Anthropic will introduce new headers or flags that will be used to disrespect non-approved requests in future versions (until the next leak, this already happened once already), and they've been squeezing third party harnesses (opencode) by default. Great lessons