Anthropic Faces Claude Code Leak: An Embarrassing Setback

This week, some of Anthropic's secrets were exposed, giving competitors a glimpse into the workings of its popular AI agent, Claude Code. Anthropic accidentally leaked the source code of its AI agent, which quickly circulated on the Internet, attracting millions of views and adaptations on GitHub. To control the spread, Anthropic sent a takedown request for copyright infringement.

When a segment of the source code of Anthropic's renowned AI agent, Claude Code, was published on GitHub on Tuesday, it triggered a real rush. Engineers from all backgrounds rushed to explore it, hoping to glean insights and perhaps use it to improve their own projects. If relying on content created by others to enhance intelligence sounds familiar, it's because this is exactly what major AI companies have been doing for years in their competition to train their language models — Anthropic included. Thus, it was not without a certain irony that Anthropic quickly issued a takedown notice to prevent engineers from accessing the leaked code.

“We issued a DMCA takedown against a repository hosting the source code of Claude Code and its forks,” said a spokesperson for Anthropic, referring to the Digital Millennium Copyright Act.

Anthropic, OpenAI, and Google have all faced lawsuits regarding their use of copyrighted material — including published books, articles, scientific journals, and other content found online — without explicit permission. In response, authors, artists, and publishers have used copyright law to hold them accountable and often seek payments.

In September, a court ordered Anthropic to pay $1.5 billion in damages as part of a class action lawsuit brought by authors and publishers — including lead plaintiffs Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson — for allegations that it had used pirated books and underground libraries to train Claude.

Reddit sued Anthropic last June for extracting volumes of user-generated content to train its models without permission or compensation from users. Last month, Universal Music Group, Concord, and ABKCO also filed a lawsuit against Anthropic for illegally uploading over 20,000 copyrighted songs, again with the aim of training its models.

Now, the roles are reversed, and Anthropic is relying on copyright laws to protect its own creations. “We are putting measures in place to prevent this from happening again,” said a spokesperson for Anthropic.

Fortunately for the company, the leak may not be as severe as some thought. Paul Price, a cybersecurity specialist and founder of the ethical hacking firm Code Wall — which recently discovered vulnerabilities in McKinsey's internal chatbot, Lilli — stated that the Anthropic leak did not expose any critical elements.

“It’s more embarrassing than harmful. Most of the really interesting elements are in their internal models, and that hasn’t been leaked,” he told Business Insider. He clarified that the company had inadvertently exposed its "harness" — a software infrastructure typically used to connect large language models to the broader context in which they are used.

“Claude Code is one of the best agent connection systems out there, and now we can see how they tackle tough problems,” Price added, noting that this could also provide useful insights to competitors.

The leak also highlighted a paradox of the AI hype cycle: the same tools that make the creation and deployment of products faster than ever also facilitate the leaking, replication, and instant dissemination of information — sensitive or not.