Anthropic has confirmed that the recent leak of 512,000 lines of source code for its AI coding assistant Claude was the result of human error, not a malicious security breach, according to the Times of India. The company disclosed that on March 31, a 59.8 MB source map file was accidentally released, exposing extensive TypeScript code that included security validators and permission models, as reported by VentureBeat.
The leak involved nearly 2,000 internal files, according to Guardian World, and revealed future features Anthropic had been developing, such as Kairos, a persistent background system designed to assist users across sessions, according to Ars Technica. This unintentional exposure has prompted the company to initiate widespread DMCA takedown notices to remove the code from the internet, although this action mistakenly pulled down thousands of legitimate GitHub forks, which GitHub later restored, as Ars Technica also details.
Security experts emphasize that the exposure of sensitive code presents significant risks, including revealing attack surfaces and security weaknesses, leading to industry calls for enterprises to reassess their AI security protocols, according to VentureBeat. Additionally, the rapid rewriting of the leaked code by an engineer using AI tools to bypass copyright restrictions, reported by the Times of India, highlights the challenges in maintaining control over proprietary AI developments.
Looking ahead, Anthropic’s response to tighten internal security measures and the tech community's handling of leaked AI source code will be critical to safeguarding future innovations. Observers will be watching closely for any further unauthorized disclosures and how vendors balance transparency with protection of AI intellectual property.
Claude
Anthropic
GitHub
