Home / Technology / AI Leaks Passwords: "Don't Ask Again" Flaw Exposed
AI Leaks Passwords: "Don't Ask Again" Flaw Exposed
9 Nov
Summary
- An AI's "don't ask again" setting inadvertently shared sensitive data.
- Thousands of secrets, including passwords and API keys, were exposed on GitHub.
- The security flaw was reported as far back as August 2025 but remained unaddressed.
A critical security vulnerability in Claude Code has led to the accidental exposure of thousands of sensitive secrets on public GitHub repositories. Researchers discovered database passwords, API tokens, and root logins embedded within Claude Code settings files. The leak arises from the AI's "Yes, and don't ask again" approval for shell commands. This choice is saved in a local settings file, which can sometimes be coded as a pattern, or the command itself is written as is, potentially including sensitive information.
While the `.claude/settings.local.json` file is usually added to a developer's global .gitignore, this setting is not repo-level. This allows the file to be uploaded publicly if a repository is cloned or a folder becomes a repo. Apiiro researchers found over 100,000 instances of this settings file on public GitHub, with approximately 1,100 containing secrets from 63 organizations.
This issue was first flagged in a Claude Code bug report in August 2025, but reports were closed due to inactivity. Developers are now investigating potential solutions, with recommendations including using Claude Code's auto mode or employing secret scanners. However, even some popular scanners failed to detect certain exposed secrets, underscoring the ongoing challenges in securing AI systems.
