Claude Code saved passwords from commands: secrets found in 1.1% of settings files on GitHub
Click "don't ask again" in Claude Code on a command that contains a password, and the entire command gets written to .claude/settings.local.json. Apiiro found 104 185 such files on GitHub, and 1 122 of them (1.1%) held a secret. Version 2.1.323 closes the hole, but the update does not revoke keys that have already leaked. Update, then rotate everything that ended up in the file.
AI-processed from Habr AI; edited by Hamidun News
Picture this: you ask Claude Code to check staging, the agent proposes running psql with the password right in the command, and for the tenth time you click "Yes, and don't ask again." The whole command, password included, is written to .claude/settings.
local.json inside the project. If that file later lands in git, the password is exposed publicly.
That is how, according to Apiiro, 104 185 copies of this file ended up on public GitHub, and 1 122 of them (about 1.1%) contained at least one password, token or key. Anthropic has already shipped a fix in version 2.
1.323, but it does not undo the damage: anything that has already leaked will have to be rotated by hand. So it is worth opening this file in your projects right now.
How a password ends up in a project file
In manual mode, Claude Code asks for permission on almost every shell command. If you choose "don't ask again," the rule is saved to .claude/settings.local.json in the repository root and applies in all later sessions. The rule copies whatever was in the prompt. Sometimes that is a harmless pattern such as "npm test *". Sometimes it is one specific command in full, for example a psql call with the password in a PGPASSWORD variable. In that case the password sits in the file in plain text.
The scenario is mundane. You have already approved a safe command ten times, you are tired, and you click "don't ask again" on the eleventh, where the agent checks staging using a password from a local config. A secret that used to live in a secrets manager or an environment variable now also sits in a JSON file inside the project.
According to The Hacker News, citing Apiiro, the 104 185 copies of the file were found in 103 617 repositories. 1 122 copies contained at least one secret. In total there were 1 379 secrets belonging to 63 organizations. More than a third give direct access to databases, including 379 database passwords. Researchers also found 43 SSH logins, 42 GitHub, GitLab and npm tokens, and 23 AWS keys. Ars Technica rounds the same data to "more than 100,000" files and "about 1,100" with secrets. There is no contradiction here.
An important caveat: the researchers did not check whether the secrets they found are still valid, and they do not claim anyone has already used them. One of the notified companies confirmed that its key was live. This is an exposure risk, not a proven breach. Also, Apiiro sells security tools, which is worth keeping in mind when judging the scale.
Why a "local" file ended up in someone else's clone
The name settings.local.json sounds safe. When Claude Code creates the file, it adds the path to the global git excludes, but only on that computer. No entry appears in the project's own .gitignore. On a new laptop, in CI, on a colleague's machine after cloning, or in a folder that became a repository later, that protection may be missing. And if the file has already been added to the git index, ignoring it does not help at all.
Boris Cherny, head of Claude Code, wrote on Hacker News that the file is supposed to stay local and that in such cases it was created and committed manually. But according to Apiiro, in 539 of 1 105 traced repositories the file was changed by a commit that lists Claude as co-author. And Anthropic's own changelog acknowledges the new-machine scenario and the "folder that became a repository later." Blaming it all on careless users would be inaccurate.
Standard scanners fall short too. Gitleaks found a secret in 564 of the 1 122 files, TruffleHog in 643, and 215 files were missed by both. The reason is simple: scanners are good at catching known key formats, while a password after PGPASSWORD= or sshpass -p looks like an ordinary word.
What changed and what to check
In Claude Code 2.1.323 (November 12), "don't ask again" no longer saves rules that look like a secret, or rules that would land in a file git already tracks. Such an approval applies only to the current session. The release also fixes the file's ignore handling on a new machine and in repositories created later. The update does not delete old rules, clean git history or revoke keys.
What a developer should do:
- Check your version with claude --version and update to at least 2.1.323.
- Open .claude/settings.local.json in every project and delete rules containing passwords, tokens and connection strings.
- Approve patterns like "Bash(psql *)", and in the command itself reference a variable such as $DATABASE_URL rather than the value.
- Search for the file in history: git log --all --full-history -- .claude/settings.local.json.
- If a secret was ever there, rotate it first, and only then run git rm --cached and add the file to the repository's .gitignore.
There is no need to scrub the shared .claude/settings.json indiscriminately: it is committed on purpose. But it is worth checking too, since according to Apiiro, secrets turned up in 298 of 76 011 public copies.
For teams: add settings.local.json to the organization's .gitignore template, search for the file across all repositories along with their history, account for forks, and configure scanners to catch passwords inside commands. The .cursor and .gemini folders and agent logs can also store secrets.
The takeaway is practical. We trust coding agents with more and more actions, and their settings files are becoming a new place where secrets accumulate, much as .env once was. The update lowers the risk of new leaks, but old ones are closed only by rotating keys.
- Nov 18, 2026TokenEconomizer plugin cuts AI coding agent costs by 18.4–30.9% without changing the model
- Nov 18, 2026Five AI Models Tried to Fix the Same Bug: Two Succeeded, Three Broke More Than They Fixed
- Nov 18, 2026Qwen on two DGX Spark units runs 34–59% faster in tokens per second and outpaces GLM
- Nov 18, 2026Developer gives a web AI chat hands: an MCP server with 14 tools and 21,000 lines of code
Key Questions Answered
How a password ends up in a project file?
In manual mode, Claude Code asks for permission on almost every shell command. If you choose "don't ask again," the rule is saved to .claude/settings.local.json in the repository root and applies in all later sessions. The rule copies whatever was in the prompt. Sometimes that is a harmless pattern such as "npm test ". Sometimes it is one specific command in full, for example a psql call with the password in a PGPASSWORD variable. In that case the password sits in the file in plain text. The scenario is mundane.
Why a "local" file ended up in someone else's clone?
The name settings.local.json sounds safe. When Claude Code creates the file, it adds the path to the global git excludes, but only on that computer. No entry appears in the project's own .gitignore. On a new laptop, in CI, on a colleague's machine after cloning, or in a folder that became a repository later, that protection may be missing. And if the file has already been added to the git index, ignoring it does not help at all.
What changed and what to check?
In Claude Code 2.1.323 (November 12), "don't ask again" no longer saves rules that look like a secret, or rules that would land in a file git already tracks. Such an approval applies only to the current session. The release also fixes the file's ignore handling on a new machine and in repositories created later. The update does not delete old rules, clean git history or revoke keys. What a developer should do: Check your version with claude --version and update to at least 2.1.323.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.