Anthropic Accidently Leaks Code Behind Its Hit Tool

Anthropic just got a crash course in how fast its own AI buzz can backfire. The company behind Claude Code, an AI coding assistant with "viral popularity" among developers, is scrambling after a Tuesday update briefly exposed the internal instructions that tell the tool how to behave—its so-called "harness"—on GitHub, per the Wall Street Journal. By Wednesday, Anthropic had pushed the platform to remove more than 8,000 copies and derivatives via copyright claims. The company says only some internal source code leaked due to "human error," with model "weights" still protected. Still, the exposed material, reportedly some 500,000 lines of code, offers rivals and hobbyists a detailed blueprint for recreating Claude Code's behavior—and gives potential hackers new angles to probe.

Developers poring over the files highlighted a "dreaming" system for consolidating memory, instructions for sometimes masking that the agent is an AI when posting code, and even future product plans. Though Anthropic is moving to contain the damage as it eyes an IPO and a sky-high valuation, at least one programmer has already released a reimplemented version designed to sidestep takedowns. The result is that "legions of startups and developers now have a detailed road map to clone Claude Code's features without needing to reverse engineer them," per the Journal, while hackers have new ways to probe for vulnerabilities, creating risks for users. This is the second time source material for Claude Code has been exposed in just over a year, per Axios.