Anthropic Leak Reveals Code for Claude

“Anthropic is racing to contain the fallout after accidentally exposing the underlying instructions it uses to direct Claude Code, the artificial-intelligence agent app that has won the company an edge with developers and businesses.

 

By Wednesday morning, Anthropic representatives had used a copyright takedown request to force the removal of more than 8,000 copies and adaptations of the raw Claude Code instructions -- known as source code -- that developers shared on programming platform GitHub. It later narrowed its takedown request to cover just 96 copies and adaptations, saying its initial ask reached more GitHub accounts than intended.

 

The leak of "some internal source code" didn't expose customer information or data, a spokesman for Anthropic said. Nor did it divulge the valuable inner mathematics -- sometimes called weights -- of the company's expensive and powerful AI models.

 

"This was a release packaging issue caused by human error, not a security breach. We're rolling out measures to prevent this from happening again," the spokesman said.

 

But the leak did reveal commercially sensitive information, including Anthropic's proprietary techniques, tools and instructions for cajoling its AI models to work as coding agents. Those techniques and tools are called a harness because they are what allow users to control and direct those models, much like a harness allows a rider to guide a horse.

 

The result is that Anthropic's competitors and legions of startups and developers now have a detailed road map to clone Claude Code's features without needing to reverse engineer them -- something that is already common in the cutthroat AI race.

 

The leak also gives hackers a large amount of new information to probe for bugs they could use to exploit the Claude Code software, or manipulate its Claude AI model into helping with their cyberattacks, creating risks for Anthropic and the developers who use its tools.

 

The leak is a blow for Anthropic because it risks both undermining its reputation for safety and revealing valuable trade secrets in the pitched battle for enterprise customers. Anthropic has been riding a wave of growing use because of the viral popularity of Claude Code, helping it close a new round of funding that values the company at $380 billion ahead of a possible public offering this year.

 

Much of the excitement about Claude Code is about how it manages to stitch together the company's AI models and coax them into working well in a way that helps developers get work done -- something called "tooling" that in AI is as much an art as a science.

 

The sensitive Claude Code information was inadvertently disclosed on Tuesday when the company updated the AI tool. Like most proprietary software, Claude's source code is usually obfuscated and hard to reverse engineer.

 

Except this time, the company posted to GitHub a type of file that linked back to the source code that outsiders could download and interpret.

 

A user on the social-media platform X quickly noticed the leak and spread the word. Within hours, copies were multiplying, leading to a game of cat-and-mouse.

 

Programmers combing through the source code so far have marveled on social media at some of Anthropic's tricks for getting its Claude AI models to operate as Claude Code. One feature asks the models to go back periodically through tasks and consolidate their memories -- a process it calls dreaming.

 

Another appears to instruct Claude Code in some cases to go "undercover" and not reveal that it is an AI when publishing code to platforms like GitHub. Others found tags in the code that appeared pointed at future product releases.

 

The code even included a Tamagotchi-style pet called "Buddy" with which users could interact.

 

After Anthropic requested that GitHub remove copies of its proprietary code, another programmer used other AI tools to rewrite the Claude Code functionality in other programming languages.

 

Writing on GitHub, the programmer said the effort was aimed at keeping the information available without risking a takedown. That new version has itself become popular on the programming platform.

 

The leak is useful because it names hidden features and coming models, but it is unlikely to be of use to hackers, said Dan Guido, chief executive of cybersecurity firm Trail of Bits.

 

Hackers already could reverse engineer the code before the leak, while Claude Code is frequently rewritten meaning the leak will soon be obsolete, he said. "The leak is embarrassing but not dangerous," Guido said.” [1]

 

1. Anthropic Leak Reveals Code for Claude. Schechner, Sam; McMillan, Robert.  Wall Street Journal, Eastern edition; New York, N.Y.. 02 Apr 2026: A1.