A button that runs your tools is also a button that can do damage. Today, cyber security, the calm version. You will learn the few guards that let you automate without handing a stranger the keys. Boring on purpose. Boring is what keeps you out of the news.

Step one: never write a password or API key into a workflow file. Put secrets in environment variables and read them by name. If the file leaks, it leaks instructions, not your accounts. This single habit prevents most of the disasters you have read about.

Step two: when you set up an MCP server, give it the least access that still works. A read-only key for a tool that only reads. One folder, not the whole drive. If the agent ever does something dumb, the blast stays in one small room you chose on purpose.

Step three: put a human gate on anything that sends, deletes, or pays. The agent stops, shows the exact action, and waits for your yes. Speed is nice, but a forced pause before the irreversible step is the cheapest insurance you will ever buy.

Step four: have every run write a plain log, what it touched and when. If something looks off, you read the log instead of guessing. An honest record is half of security. You cannot fix what you cannot see, and you cannot prove what you did not write down.

Step five: turn the agent on yourself. Ask Claude Code to scan your config for keys left in files or tools with too much access. It finds the holes faster than you will, because it never gets bored reading the boring lines.

Proof, plainly: in our own runs, nothing posts or deletes without a gate and a log, and we keep keys out of every file we show you. The public walkthrough still has the brakes on. We are not braver than you. We are just careful in the open.

So: keys in the environment, smallest access per tool, a gate before anything irreversible, a log of everything, and an audit you ask the agent to run. Do these five and you can automate without losing sleep. The checklist is at office dot temerarii dot xyz.