Anthropic's Claude Code caused a large-scale leak of confidential data. Due to a flawed settings-saving mechanism, thousands of database passwords, API tokens, and access keys ended up publicly exposed on GitHub.
Cybersecurity specialists at Apiiro reported the serious vulnerability after discovering the issue while analyzing public repositories, Ars Technica reports. According to their data, the leak is related to how the tool handles user permissions to execute automated commands in the terminal.
When a developer allows the algorithm to execute a command and selects the "Yes, don't ask again" option, Claude Code saves that choice in a local .claude/settings.local.json file. The problem is that, along with the rule, the command itself is written there in its original form, including any passwords or tokens it contained.
Anthropic claims it adds this file to the global ignore list (.gitignore) to prevent it from being published. However, this setting does not always work at the level of individual repositories. If the project is cloned to another machine or an ordinary folder later becomes a repository, the file with passwords may be uploaded to the internet. Moreover, sometimes Claude Code itself automatically commits these settings without direct instruction from the user.
The scale of the problem turned out to be significant. Apiiro researchers found more than 100,000 such configuration files on GitHub. About 1,100 of them, belonging to 63 different organizations, contained critical secrets such as public API tokens and private keys. Even after files are deleted from the main repository, the data may remain in its copies (forks) if the developer does not exercise due caution.
The first reports of this bug appeared back in August 2025, but they were closed at the time due to inactivity. Currently, Claude Code team member Boris Cherny confirmed on Hacker News that the developers are looking for a solution. For now, he advises programmers to switch to an automated mode of operation and use special secret scanners to clean the code. However, cybersecurity experts note that popular scanners such as gitleaks and TruffleHog were unable to detect more than two hundred of the passwords found.
The day before, Anthropic updated its line of neural networks, unveiling the Claude Fable 5.2 and Claude Mythos 5.2 models. They are aimed at solving the most complex technical tasks and during testing were able to independently design electronic devices and write an operating system from scratch.