Wednesday, November 18, 2026

Top 5 This Week

Related Posts

Claude Code: Permissions file exposed more than 1,300 secrets on GitHub


Can a simple convenience option expose a company’s passwords? This is what a study by the security company shows Apiiroreported on November 7, 2026 by the specialized site The Hacker News. Researchers found on GitHub more than 1,100 public files from Claude Code, Anthropic’s programming assistant, containing database passwords, cloud access keys or server identifiers. At issue: the way the tool memorizes the commands that the user has authorized once and for all.

Where does the problem come from?

Claude Code is an agent capable of executing commands on the developer’s computer. Before launching one, he asks for authorization. User can respond “Yes, and don’t ask again” (“Yes, and don’t ask again”). According to the official Anthropic documentation, Claude Code then records this approval in the form of an authorization rule in a local settings file, .claude/settings.local.jsonplaced at the root of the project.

A rule can be a pattern, such as “allow all test commands”, or an exact command. In this second case, according to the researchers, the order is recorded word for word. If a password or key appeared in plain text, for example to connect to a database, it is copied in readable text into the file.

A file excluded from git, but only on one computer

Anthropic has provided protection. Its documentation indicates that the first time Claude Code writes this file to a git repository, he adds it to the git global exclude file of the user, so that it does not leave in the commits. But this protection only applies to the computer concerned. According to the researchers, Claude Code never adds the file to the .gitignore of the project itself, the file shared with the whole team. A colleague, another position or a file created by hand can therefore send it to GitHub.

The researchers noted another phenomenon: in 539 of the 1,105 repositories whose history they traced, a commit signed with Claude as co-author modified this file. They deduce that the agent itself often added it to a commit.

What the researchers found

Apiiro searched the public GitHub repositories for this file. The main findings, as reported by The Hacker News:

Indicator Result
Public copies of file found 104,185, in 103,617 deposits
Copies containing at least one secret 1,122, or approximately 1.1%
Different secrets in these copies 1,379
Database passwords 379
Total, including history and other files 2,539 secrets in 1,994 repositories

Among the secrets also include 43 server connection identifiers, 42 access tokens to GitHub, GitLab or npm and 23 Amazon Web Services keys. The files belong in particular to 63 organizations, including, according to the researchers, a database publisher and the owner of a medical records base. In half of the affected repositories, the file had been online for at least 217 days.

The researchers specify that they have not tested whether these secrets still work. A notified company, however, confirmed that its key was active.

Detection tools that miss

The study points to another weakness: common secret scanners missed part of it. Gitleaks flagged 564 of the 1,122 files and TruffleHog flagged 643, and neither spotted 215 of them. Most of the omissions concern passwords slipped in the middle of an orderwhich these tools, designed to recognize key formats, detect poorly.

A problem specific to Claude Code?

Out of 16 programming agents whose documentation the researchers compared, Claude Code would be the only one confirmed to save the complete approved command in a project file by default. The phenomenon is not new, however: a report submitted to Claude Code’s public repository in August 2025 already described this behavior. According to The Hacker News, it was automatically closed by a bot after 60 days of no activity, like the four related reports cited by the researchers, with no response from Anthropic. The article does not report any announced correction or any statement from the company.

This is not the first security alert around AI programming tools: in October 2026, a booby-trapped extension targeting users of Claude Code and Cursor had already shown that these assistants have become a target.

What to do if you use Claude Code?

The researchers propose several measures, to be applied in this order:

  • to research .claude/settings.local.json and .claude/settings.json files in all organization repositories, including history;
  • change immediately any password, token or key they contained, even before cleaning the repository, because a published secret must be considered compromised;
  • remove the file from tracking from git and add it to the project’s .gitignore;
  • DELETE local rules that contain a secret, even if they have never been published;
  • never type a secret in a command: use environment variables powered by a secrets manager;
  • approve models of commands, for example all commands in a test script, rather than exact commands.

For security teams, researchers advise treating agent configuration folders (.claude, .cursor, .gemini) as sensitive files, including them in automatic checks before each release, and adding the local Claude Code file to the organization’s .gitignore templates.

What about businesses in Europe?

When personal data is accessed by an exposed secret, such as a customer or patient database, the incident may constitute a data breach under the GDPR. The data controller must then document it and, if it presents a risk for individuals, notify it to the data protection authority, the CNIL in France, within 72 hours of becoming aware of it.

Programming agents, a new attack surface

The case illustrates a broader risk: programming agents accumulate information in their configuration, memory or history files that no one rereads, and which ends up circulating with the code. As their use becomes more widespread in businesses, the management of these files becomes a cybersecurity subject in its own right. Developers who use these tools on a daily basis should check what their configuration files contain now. A response from Anthropic, whether hiding the secrets or adding the file to the project’s .gitignore, would reduce the risk at the source.

Shredder Smith
Shredder Smith
Shredder Smith is the lead curator and digital persona behind topaitools4you.com, an AI directory dedicated to "shredding" through industry hype to identify high-utility software for everyday users. Smith positions himself as a blunt, no-nonsense reviewer who vets thousands of emerging applications to filter out overpriced "wrappers" in favor of tools that offer genuine ROI and practical productivity. The site serves as a watchdog for the AI gold rush, providing categorized rankings and transparent reviews designed to help small businesses and creators navigate the crowded tech landscape without wasting money on low-value tools.

LEAVE A REPLY

Please enter your comment!
Please enter your name here

Popular Articles