Claude Code’s “Don’t Ask Again” leaked secrets in 1,122 public repos

The click that silences Claude’s permission prompts quietly writes your passwords to disk — and the same week, the agent was caught mailing user emails to every site it visits.

Nowline NOV 8 3:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • 2,539 live secrets, sitting in Claude’s own approval files

    Apiiro scanned public GitHub and found 1,122 `.claude/settings.local.json` files carrying 1,379 secrets — 2,539 once you count older versions, the shared `.claude/settings.json` and saved chat logs, across 1,994 repos. The cause: clicking “Yes, and don’t ask again” saves your exact command — plaintext DB passwords, tokens and keys included — straight into the file. Of 16 coding agents tested it’s the only one that does this by default, reproduced on Opus 5.5 and Fable 5.1. The haul: 379 database passwords, 23 AWS keys, 42 GitHub/GitLab/npm tokens, 43 SSH logins. One company confirmed its leaked key was still live.

  • Your .gitignore didn’t save you — here’s what will

    Claude Code writes the ignore rule to your *global* git excludes, not the project’s `.gitignore`, so the file rides into any Claude-co-authored commit and onto every teammate’s machine — 539 of 1,105 traced repos were committed exactly that way. Gitleaks and TruffleHog missed 215 of the files because the secrets are typed inside commands, not stored as recognizable keys. Do this today: rotate every credential these files ever held, `git rm --cached` them into the repo’s `.gitignore`, approve command *patterns* like `Bash(npm run:*)` instead of single commands, and reference `$DB_URL` from a secrets manager rather than pasting literals.

  • And Claude Code was mailing your address to strangers

    A separate open issue: Claude Code stuffed your account email into the `User-Agent` header of its `curl` and web-search calls, so every site it visited harvested it — the reporter says it drew dozens of scam emails a day. The `cc-plugin-privacy` mod (on by default since 2.1.317) now prompts before sending the email, but maintainers admit it misses cases where Claude writes and runs its own script; full removal from the model’s context is slated for a release this week. Until then, update past 2.1.317 and watch what your agent puts in outbound headers.