Anthropic is currently dealing with an accidental leak of Claude Code’s source code. Even though the company has termed this incident a ‘human error’, the leak may have reportedly revealed commercially sensitive information about the AI coding agent that recently wiped out trillions from the stock market.According to a Wall Street Journal report, the source code leak includes Anthropic’s proprietary techniques, tools, and instructions for directing its AI models to act as coding agents. These techniques and tools are collectively referred to as a “harness,” a term that reflects how they allow users to control and guide the models, just as a harness allows a rider to direct a horse.As a result, Anthropic’s competitors, as well as many startups and developers, now have a clearer path to copying Claude Code’s features without having to reverse-engineer them, which is already common in the AI space.The leak also provides hackers with additional information to search for vulnerabilities that could be exploited to compromise the Claude Code software or influence its AI model to assist in cyberattacks, posing risks for Anthropic and developers who rely on its tools.
How this leak can be a huge blow for Anthropic
The incident poses a risk to Anthropic on two fronts: its standing as a safety-focused AI company, and the exposure of trade secrets at a time when competition for enterprise customers is intensifying. Claude Code’s growing adoption among developers had helped Anthropic close a new funding round valuing the company at $380 billion, ahead of a possible public offering this year.A significant part of Claude Code’s appeal lies in how it connects the company’s AI models and guides them to work in ways that help developers complete tasks, an approach known as “tooling” that practitioners consider as much craft as technical execution.Earlier this week, Anthropic disclosed sensitive Claude Code information during an update by mistake. Instead of keeping the source code complex and confusing, the company uploaded a file to GitHub that linked to code accessible and interpretable by outsiders.An X user quickly realised this leak shortly after and made it publicly known. Within hours, this code began circulating across various platforms, sparking discussion among programmers. As programmers reviewed this information, they have been able to point out some of its characteristics, such as its “dreaming” feature for task organisation, how to “use it while undercover,” possible future updates, and an interactive feature called “Buddy.”Within a day, Anthropic issued copyright takedown requests, leading to the removal of more than 8,000 copies and adaptations of the code from GitHub.The report went on to claim that some developers tried to ensure that access remained possible, even after all these efforts. One developer used AI tools to copy Claude Code’s work into other programming languages and post it on GitHub. They did this so that people could still access it, and attempts to take it down would be stopped. The recreated work has become popular on the site.Meanwhile, Anthropic noted that the leak involved “some internal source code” but did not expose customer data or the underlying model weights. “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,” a spokesperson told WSJ.
