Anthropic source code leak: Claude Code files exposed in second security lapse within days and ahead Mythos launch

Anthropic

• Anthropic leaked 500,000 lines of Claude Code source code
• No customer data or credentials were compromised in the leak
• Leak may expose Claude Code’s internal design to rivals

Did our AI summary help?

Anthropic has confirmed that internal source code related to its coding tool Claude Code was unintentionally made public, marking the second disclosure involving the company within a week. The incident follows an earlier exposure that revealed details about an unreleased AI model internally referred to as Mythos.

What was leaked

According to Fortune's report, the latest incident exposed close to 500,000 lines of code across around 1,900 files linked to Claude Code. The company said the issue was caused by a release packaging error during deployment, adding that no customer data or credentials were compromised.

Claude Code is a key product for Anthropic, widely used by enterprises. The leaked files reportedly include parts of the software layer that manages how the AI model interacts with tools, systems, and instructions. This layer, often referred to as an agentic harness, governs execution flows and safety mechanisms.

How the leak happened

The exposure appears to have occurred when the company uploaded the full source package to a developer platform instead of a compiled version. This resulted in public access to internal code that is typically restricted.

Anthropic stated that the issue was due to human error and not a breach of its infrastructure. The company also said it is implementing measures to avoid similar incidents in future releases.

Security and competitive concerns

Cybersecurity researchers suggest the leak could allow developers or competitors to study internal architecture and processes. While the core AI model weights were not exposed, the surrounding system design may provide insights into how the tool operates and enforces safeguards.

There are also concerns that such information could help advanced actors understand system behavior, including APIs and deployment patterns, potentially aiding attempts to bypass protections.

The leak adds to scrutiny around AI system security, especially as models grow more capable and are deployed in enterprise and development environments.