Quick win: hand your AI a secret it never sees

AI agents now do real work for us: deploy apps, run servers, send email. Sooner or later one of them needs a password or an access token. The question is how you hand it over.

The quick win

My agent and I worked out a little dance. It feels odd the first time and natural by the third.

1. The agent gives me a one-line command to run. I paste it into my terminal and don't hit Enter.
2. I go to wherever the secret lives (my password manager, the website that issued it) and copy it.
3. I come back and hit Enter.

The command reads the secret straight off my clipboard, delivers it to the machine that needs it, clears the clipboard, and reports one thing back: "Set (108 characters)." The agent learns that it worked. It never learns the secret. The secret never lands in the chat, on screen, or in a log someone could read later.

That's the whole trick: the AI writes the delivery truck but never opens the box.

The top five ways people pass secrets instead (please don't)

1. Pasting it into the chat. "Here's my OpenAI key, go wild." Your key now sits in a conversation log on someone else's servers, in your chat history, and maybe in training or review pipelines, depending on your settings. You wouldn't read your bank password out to a call center that records every call. Same thing.

2. Saving it in the code. The AI helpfully puts your key in a settings file, the file gets saved to Git, and Git remembers it forever. Deleting it later doesn't help, because the history still has it. Then the project gets copied, shared and forked, and the key travels with every copy. Bots scan public code for exactly these keys, around the clock, and find them in minutes.

3. Email and Slack. I know the head of IT at a company I won't name who sends full admin passwords to anyone who asks, in a single email or Slack message. Email gets forwarded. Slack gets searched, exported and screenshotted, and gets new members every week. "It's private" means "it's private until it isn't." Use a password manager's share feature, or a link that self-destructs after one view.

4. Leaving it lying around in plain sight. On the screen during a demo, in a screenshot, on a sticky note, in a shared doc called "passwords (DO NOT SHARE)". Or my own slip this week: I pasted the token one time too many during the dance above, and it landed on the command line, where the AI's log picked it up. The trick only works if you follow the steps. When a secret leaks, don't debate it. Make a new one and switch it out.

5. Using someone else's secrets. This is the one that ends careers. It isn't your key you're pasting into a free AI tool; it's your company's database login, or your customer's. Now customer records (health data, financial data, personal details) have flowed somewhere they were never allowed to go. That's the stuff of HIPAA, GDPR and SOC 2 reports, and of very bad weeks. If the data isn't yours, the decision about where it goes isn't yours either.

The short version

• Give the AI a way to deliver a secret, never the secret itself.
• Keep secrets in a password manager, never in chat, code, email or Slack.
• When one leaks, replace it. Don't hope.

---

For the engineers: the delivery script is about ten lines. It does pbpaste, pulls out the token with a regex, pipes it over ssh on stdin (so it never appears in the process list or shell history), writes it to a chmod 600 env file, prints only its length, then pbcopy </dev/null. Have your agent write one for each place a secret needs to go.