Apiiro published a report on 7 November 2026 saying that Claude Code’s “Don’t Ask Again” option exposed passwords in 1,122 public GitHub files. The post drew more than 300 points on Hacker News. I could not load the full article, so I am working from the headline and the summary only. I am not going to guess at the technical detail beyond that.

The excitement is about speed

Most of the talk around coding agents is about how fast they build. You describe what you want and a working tool appears. That part is real. Vibe coding collapses the distance between an idea and working software, and I use it to replace many tools and scripts with ones I build myself.

The report points at what that speed skips. “Don’t Ask Again” exists to save you from approving every step. That is a reasonable thing to want. It also means one fewer moment where a person looks at what is about to happen.

Is there a real need here? Yes, and it is a narrow one

The need behind the convenience is genuine. Nobody wants to click “approve” two hundred times in a session. The question is what changed when the clicks went away.

The build got faster. The check on what leaves your machine did not get any smarter. A prompt like “build me a booking tool and put it on GitHub” says nothing about which files hold passwords. The tool does not know your business. It does not know that one file is harmless and another one opens your database. Only you have that context.

That is the gap I keep coming back to with vibe coding. Every build still needs back-and-forth on alignment, strategy, approach and security. AI cannot supply that without a human giving it context and experience. You do not get it from a single prompt.

What this looks like for a normal operator

Take a clinic owner who builds a small tool to manage appointment reminders. It connects to a messaging account and a patient spreadsheet, so it needs passwords and keys to work. The owner is not a security person and does not need to be. They built it for their own use, which is where vibe coding really shines.

Then the code gets pushed to a public repository, perhaps because that was the default or because the agent offered to set it up. If a password sits in one of those files, it is now public. The owner never decided to publish a password. Nobody made that decision at all, and that is the problem.

I would not frame this as the AI being reckless. The same is true of a human developer in a hurry. What matters is where the rule lives. If the only protection is the AI’s judgement about which files are sensitive, then the protection is only as good as its judgement on the day.

Rules belong in the system, not in the prompt

My rule is to use rules where being wrong is costly and AI where judgement is needed. Publishing a password is costly. So the rule should be enforced by the system around the AI, and should not depend on the AI remembering to be careful.

I apply the same thinking to AeroChat. A merchant sets how much it can refund, and the platform holds that limit. Anything above it goes to the merchant’s team with the order details. The limit does not rely on the AI choosing to behave. Code you build for yourself deserves the same treatment.

What a business should do

If you build your own tools with a coding agent, set these up before the next project, not after.

  • Keep passwords and keys in a separate settings file that is excluded from the repository, and never inside the code itself.
  • Make new repositories private by default. Decide on purpose when something becomes public.
  • Turn on the secret scanning and push protection that your code host offers, so a push containing a password is blocked by the platform.
  • Have a person read the list of files before anything is published. This takes two minutes and needs no technical skill. Ask “which of these could open something?” and get an answer.
  • Treat “Don’t Ask Again” as a choice you make per task. Use it for harmless steps and keep approvals on for anything that publishes, deploys or sends.
  • If you find a password in a public file, change it straight away. Deleting the file does not remove it from the history.

Source: Apiiro: Claude Code’s “Don’t Ask Again” exposes passwords in 1,122 public GitHub files