Journal · The AI Architect · 2026-07-30
A secret is a house key — never tape it to the door. Credentials belong in secure settings, never in a message.
A secret is a house key — never tape it to the door. Credentials belong in secure settings, never in a message.
The tension isn’t whether you know that secrets shouldn’t be exposed. Everyone knows that, in the abstract. The tension is that the fastest way to get something working — a chatbot pulling from an API, a script hitting a database, a teammate debugging your pipeline at 11pm — is almost always the way that leaves a credential sitting somewhere it shouldn’t. Convenience and safety pull in opposite directions, and convenience wins by default because it’s immediate and safety only pays off later, when something goes wrong that you can’t yet picture. This is a note about closing that gap before it costs you something.
Why messages feel safe when they aren’t
When you paste an API key into a Slack message, an email, or a chat window with an AI assistant, it feels contained. You can see the whole conversation. It seems private. But messages get logged, indexed, cached, forwarded, and searched in ways that are invisible to you in the moment. A key pasted into a chat with an AI model may get stored in that provider’s logs. A key pasted into a support ticket lives in that ticket forever, readable by anyone who later gets added to the project. A key pasted into a group chat is one screenshot away from being somewhere you never intended.
The core mistake is treating “I can’t see anyone else looking” as the same thing as “no one else can look.” Those are different claims, and only the second one matters.
Try this: the next time you’re about to paste a credential anywhere — a chat, an email, a ticket, a code comment — stop and ask, “If this message were public tomorrow, would I be fine with that?” If the answer is no, the credential doesn’t go in the message. It goes in a secrets manager, an environment variable, or a vault built for exactly this purpose, and the message instead says something like “using the key from the shared vault, entry ‘prod-api’.”
The illusion of temporary
A common rationalization: “I’ll just put it here for now and delete it later.” This is the taped key on the door again — you tell yourself it’s just until you find your real key, but doors don’t check your intentions, and neither do logs. “Temporary” exposure is still exposure, because the systems that copy and store your message don’t know your plans. They just see the credential and keep it.
This matters even more with AI tools specifically, because the interface invites a conversational, low-friction mode of working. You’re troubleshooting, you’re moving fast, the assistant is right there, and it’s tempting to hand over a token just to get past a blocker. The design of the interaction — friendly, fast, forgiving — quietly erodes the caution you’d apply if you were, say, filling out a formal access request form.
Try this: build a small habit of separation. Keep credentials in one place — a password manager or secrets vault — and keep your working conversations, whether with a person or an AI, in another. When you need an assistant to use a credential, give it a reference or a scoped, revocable token meant for that purpose, not your master key. If a tool doesn’t support that separation, treat that as a signal about the tool, not a reason to break the rule.
Rotation is not a punishment, it’s a pressure valve
Even careful people slip. A key gets pasted somewhere it shouldn’t, a screenshot goes out, a demo recording captures more than intended. The fix isn’t perfection — it’s making rotation and revocation cheap and normal, so a slip is a minor event instead of a crisis.
If rotating a key requires four approvals and a change ticket, nobody will do it after a small scare, and small scares are exactly when it matters most. If revoking access takes thirty seconds, people will actually do it the moment something feels off.
Try this: pick one credential you use often and find out, right now, how long it would take you to revoke and replace it. If the answer surprises you, that’s useful information — it tells you where your real exposure sits, independent of how careful you plan to be.
None of this requires distrust of your tools or your teammates. It requires treating credentials as what they are — the thing that opens the door — and keeping them out of the ordinary flow of conversation, where they’re easy to copy, easy to forget, and easy to lose track of. The book goes further into how this plays out with AI systems specifically, where the assistant’s memory, logs, and context windows add a few more places a key can quietly end up. But the habit starts here, with a simple pause before you hit send.
Go deeper. The full method is in The AI Architect. New here? Start with the free companion pack, or explore the series.