AI company Anthropic has accidentally exposed the source code for its widely-used coding assistant Claude Code, marking the second significant data leak to affect the company in less than a week.
Fortune reports that the latest incident comes mere days after Fortune revealed that Anthropic had inadvertently made nearly 3,000 internal files publicly accessible, including a draft blog post describing an upcoming AI model called “Mythos” or “Capybara” that the company warned presents serious cybersecurity risks.
This second leak exposed approximately 500,000 lines of code contained within roughly 1,900 files. When contacted for comment, Anthropic acknowledged that “some internal source code” had been leaked as part of a “Claude Code release.” A company spokesperson stated: “No sensitive customer data or credentials were involved or exposed. 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.”
Cybersecurity experts suggest this latest leak could prove more consequential than the earlier exposure of the draft blog post. While the source code leak did not reveal the actual model weights of Claude itself, it enabled technically knowledgeable individuals to extract additional internal information from Anthropic’s codebase, according to a cybersecurity professional who reviewed the leaked materials for Fortune.
Claude Code represents one of Anthropic’s most successful products, with adoption rates climbing rapidly among large enterprise customers. The tool’s functionality derives partly from the underlying large language model and partly from what developers call an “agentic harness” — the software framework that surrounds the core AI model, directing how it interacts with other software tools and establishing crucial behavioral guardrails and operational instructions. It is precisely this agentic harness source code that has now been leaked online.
The exposure creates several competitive and security concerns. Rival companies could potentially reverse-engineer the workings of Claude Code’s agentic harness to enhance their own offerings. Additionally, some developers might attempt to build open-source alternatives based directly on the leaked code.
Hellbound and Down 