On March 31, 2026, Anthropic, the company behind the Claude artificial intelligence system, accidentally published the entire source code of a product called Claude Code inside a routine software update. A missing line in a configuration file shipped 512,000 lines of proprietary code to the public. No hack. No breach. Human error.

An important distinction: Claude Code is not the AI model itself. Claude’s underlying artificial intelligence (its “brain,” so to speak) was not exposed. Claude Code is the software application that allows developers to interact with Claude’s AI capabilities through a command-line interface. Think of it as the control panel and workflow engine that sits between the human user and the AI model. It manages how tasks are assigned, how code is reviewed, how files are edited, and how the AI’s responses are delivered. What leaked was this control layer: the tool’s architecture, its internal features, its security logic, and its product roadmap.

What happened next was unprecedented. Within hours, a developer used a competing AI system to reimagine Claude Code’s leaked source code from scratch in a different programming language, describing the result as a “clean-room Python rewrite” that captured the original’s architectural patterns without copying its proprietary source. The resulting project, called “claw-code,” became the fastest-growing repository in the history of GitHub (the dominant platform where developers host and collaborate on code), accumulating over 100,000 endorsements in a single day. Anthropic filed copyright takedown notices targeting direct copies. Over 8,000 repositories were initially removed. The AI-rewritten version survived.

The incident raises a question that every company building, buying, or depending on software should be asking: when an AI system can rewrite a proprietary codebase overnight and the result is different enough to survive a copyright takedown, what legal protections remain?

What Is a “Clean Room” and What Did the Developer Claim?

The developer who created claw-code described his project as a “clean-room” rewrite. In the software engineering community, this term refers to a process in which one team studies existing software and documents what it does, and then a separate team, which has never seen the original code, writes entirely new code from that documentation alone. The idea is that if the second team never had access to the original, whatever they produce must be independently created.

From a legal perspective, the “clean room defense” is not a formal statutory defense written into the Copyright Act. It is shorthand for an independent-creation argument. To prevail on a copyright infringement claim, a plaintiff must prove two things: that the defendant actually copied the protected work, and that the resulting product is “substantially similar” to the protectable expression in the original. A disciplined clean room process is designed to sever both links by generating contemporaneous proof that the new code was created independently. Independent creation is not merely desirable; it is a constitutional prerequisite for copyright protection itself.

This approach has a forty-year pedigree. In 1984, Phoenix Technologies used it to legally clone IBM’s PC BIOS (the basic firmware that makes a computer start up). Federal courts subsequently recognized that reverse engineering copyrighted software to understand its functional requirements can be lawful, even when it involves intermediate copying, so long as the final product does not reproduce the original’s protectable expression. The canonical case, Computer Associates v. Altai (2d Cir. 1992), upheld a rewrite where the defendant excluded the employee who had copied code, locked away the original, and used programmers who had never worked on the tainted version. The court found no copyright liability for the rewritten product, though it remanded trade secret claims for further proceedings. In Google v. Oracle (2021), the Supreme Court confirmed that functional elements of software receive narrower copyright protection, holding that Google’s reimplementation of Oracle’s Java programming interfaces was fair use.

Stated simply: when lawyers talk about a “clean room defense,” they typically mean the defendant independently reimplemented the unprotectable functional aspects of the software without copying the plaintiff’s protected expression. It is not a magic shield. It will not help if the final work still copies protected expression, and it may not defeat non-copyright claims such as trade secret misappropriation or breach of contract.

Key Legal Considerations

The claw-code incident raises several overlapping legal questions that existing doctrine does not cleanly resolve.

• The developer who created the rewrite had direct access to the leaked code. The claw-code developer did not follow the traditional clean room process of separating the team that studies the original from the team that writes the new code. Instead, the leaked source code was fed directly into an AI system, which then produced the rewrite. Whether the developer personally read the code line by line is beside the point; the AI tool that performed the reimplementation was given the actual proprietary source as its input. This raises the fundamental question of whether a process in which the creating agent (here, the AI) has direct access to the original code can qualify as “independent creation” in any legally meaningful sense. This question has not been answered by any court.
• AI-generated code may not be copyrightable, and this cuts in both directions. The U.S. Copyright Office confirmed in January 2025 that works predominantly generated by AI, without meaningful human authorship, are not eligible for copyright protection. The Supreme Court declined to review Thaler v. Perlmutter in March 2026, leaving this rule intact. For businesses whose valuations depend on proprietary software, the risk is severe: code produced by an automated process may not be copyrightable by anyone, meaning competitors can freely copy it. But this principle cuts against the original rights holder, too. Anthropic’s own lead engineer publicly stated that his recent contributions to Claude Code were entirely written by the AI tool itself. If the leaked codebase is substantially AI-authored, Anthropic may face difficulty establishing the human authorship required to enforce its own copyright.
• The legal risk may arise at the point of input, not output. Even if an AI-generated rewrite passes a substantial similarity analysis and avoids copyright infringement, the act of feeding proprietary or licensed code into an AI system may independently give rise to claims for breach of contract, violation of license terms, violation of terms of service, or trade secret misappropriation. Before inputting any code into an AI tool for reimplementation, businesses should evaluate whether doing so violates any contractual obligations, license restrictions, or confidentiality duties that govern that code. When an AI tool generates code, it does not flag or reproduce the license terms that governed the material it drew from. The output arrives without any indication that it may carry usage restrictions or attribution obligations from its source material, leaving the user unaware of potential exposure.
• Trade secret protection is a separate and independent consideration. A clean room rewrite that avoids copyright liability does not necessarily defeat a trade secret misappropriation claim. As Computer Associates v. Altai demonstrated, a court may find no copyright infringement in a rewritten product while still remanding trade secret issues. Trade secret analysis does not turn on textual similarity; it turns on whether the defendant misappropriated information the plaintiff took reasonable steps to keep secret. In a leaked-code scenario, the strength of a trade secret claim will depend on whether the owner maintained adequate security measures overall and acted promptly to mitigate the disclosure, which are highly fact-specific inquiries.
• Substantial similarity remains the central copyright question. At the end of the day, copyright infringement requires proof of substantial similarity between the original work’s protectable expression and the allegedly infringing work. If the AI-generated rewrite does not reproduce the original’s protectable expression, there may be no copyright infringement regardless of whether the AI had access to the original. However, a change in programming language does not automatically defeat substantial similarity. Courts evaluating software copyright look beyond literal text to the structure, sequence, organization, and logic of the code. Whether an AI-assisted rewrite from one programming language to another preserves enough of those structural elements to constitute infringement, or qualifies as a derivative work, is an open legal question that the current case law does not resolve.
• Traditional enforcement tools have structural limitations. Anthropic’s DMCA campaign successfully targeted direct copies of the leaked code but could not reach the AI-rewritten version, copies hosted on decentralized platforms outside U.S. jurisdiction, or code circulating in private channels. The initial takedown sweep inadvertently disabled thousands of legitimate repositories before being corrected. When a single developer can use AI to reimagine a proprietary codebase overnight in a form that may be beyond the reach of copyright takedown mechanisms, the practical enforcement calculus changes fundamentally.

Strategic Guidance

• What you generate with AI, you may not own. If competitive advantage depends on proprietary ownership of software, relying on AI-generated output introduces the risk that the output is not protected by copyright law. Unless a human author exercised meaningful creative control over the expressive elements of the code, the resulting work may not be copyrightable.
• Think carefully about what you input. Before feeding any code into an AI system, evaluate whether doing so violates license terms, contractual obligations, terms of service, or confidentiality duties. The legal exposure may arise at the moment of input, entirely independent of what the AI produces as output.
• Document everything. Maintain records of what tools were used, what inputs were provided, what human review processes were applied, and the extent of human creative contribution to the final product. This documentation serves dual purposes: it supports a chain of custody defense in any future dispute, and it establishes the human authorship necessary for copyright protection of your own AI-assisted work.
• Prepare for the enforcement gap. The Claude Code incident demonstrates that even a well-resourced company with immediate legal response capacity could not prevent the AI-assisted reimplementation of its proprietary software. Businesses should assess their own exposure and consider whether trade secret, contractual, and technical protections (not just copyright) are adequate to protect critical software assets.

Looking Ahead

No court has ruled on whether AI-assisted reimplementation qualifies as a legitimate clean room process. No court has decided whether feeding proprietary code directly into an AI tool, which then produces a rewrite in a different programming language, constitutes independent creation or produces a derivative work. These are genuinely novel questions that the existing case law does not answer, and the courts that eventually address them will be writing on a largely blank slate.

What is clear is that AI has compressed the reimplementation of complex software from a months-long engineering effort into an overnight task. The Claude Code leak demonstrates what that compression looks like in practice: 512,000 lines of proprietary code, rewritten and published to the world before the company’s lawyers could act. For businesses that depend on proprietary software, the practical implications are immediate, even if the legal answers are not.

For a deeper analysis of the intellectual property considerations raised by AI code generation, including litigation-readiness, discovery strategy, and how courts may apply existing copyright, trade secret, and contract doctrines to these novel facts, please contact Kandis Koustenis at kkoustenis@beankinney.com and Andrew Gregg at agregg@beankinney.com

Note: The Claude Code leak and claw-code rewrite are developing events as of this writing (April 8, 2026). Anthropic has not published a formal post-mortem, no litigation has been filed against the clean-room rewrite, and the legal status of claw-code remains unresolved. This analysis reflects publicly available information as of the publication date and may be supplemented as developments warrant.

Disclaimer: This analysis provides general legal insights and should not be construed as specific legal advice. Consult qualified intellectual property counsel for guidance on AI code generation and software IP compliance questions.