Apiiro, an application security company, found that a local settings file used by Claude Code, Anthropic's AI coding assistant, has leaked credentials into public GitHub repositories. When a developer approves a shell command permanently, the tool stores the full command, including any password or token in it. Apiiro counted 1,122 public copies of the file holding 1,379 distinct secrets, among them database passwords, AWS cloud keys, and logins for remote servers. The Hacker News reported the findings on November 7.
Any team using AI coding agents can commit such files by accident, and deleting them later does not erase git history. Apiiro says two common secret scanners missed 215 of the affected files. It advises rotating exposed credentials, keeping agent folders out of version control, and scanning git history. The report includes no Anthropic statement; it says related bug reports were closed automatically.
