Home / News

Claude Code's "Don't Ask Again" leaked passwords into 1,122 public GitHub files

An AI agent should handle judgment work, and a plain deterministic script or secrets manager should handle anything that touches credentials. The Claude Code password leak shows the cost of mixing the two: when an agent sees a secret and gets a standing approval, that secret can end up somewhere you never intended.

Security firm Apiiro found passwords exposed in 1,122 public GitHub files tied to Claude Code's "Don't Ask Again" permission flow. The likely mechanism: approving a command permanently saves a permission rule to a project config file. If that command had a password typed inline, the password can be saved with the rule. That file then gets committed and pushed like any other file in the repo.

What this means for small business automations

Most of the automations I build look the same: an LLM reads an invoice, a contract or an incoming lead, and regular code does the rest - writing to the CRM, calling the bank API, sending the email. The model never sees an API key. Keys live in environment variables or the client's secrets store, and the code injects them at the moment of the call. When a business owner lets a coding agent run on the same machine as production credentials and clicks "Don't Ask Again" to save time, that line disappears.

Why it matters: Run git grep for your passwords and tokens in every repo where you have used a coding agent today. Then rotate anything that shows up and add the agent's local settings files to .gitignore.

Source: Hacker News