Trust No Bot

You did not notice the moment you handed over the keys, because the moment looked like a login screen and a small button that said allow. The assistant needed access to your calendar, so you clicked the button, and it needed access to your email, so you clicked the button again, and it needed access to your documents, your browser, and the folder where the numbers live, and you clicked the button for all of them. Somewhere in that blur of clicks, a computer stopped being a tool you operate and started being an employee you supervise.

This paranoid but pragmatic guide explains, in plain language and with a healthy dose of humor, why that arrangement is a problem and what you can do about it before it matters. An AI agent is not a smarter search box, it is a program that has been given a goal, a set of tools, and permission to act on your behalf. It reads your email so it can write better drafts, opens your calendar so it can schedule your meetings, and reaches into your codebase so it can fix your bugs. Each of those actions is individually reasonable, and that is precisely the problem, because the danger rarely arrives as a villain with a plan, and the danger arrives as a very capable assistant that was tricked, confused, or over-permissioned.

This book covers the paranoia starter kit, the anatomy of an agent, the art of prompt injection, the ghost in the inbox, why your keys are not your keys, the least privilege love story, the oversight dilemma, the human in the loop, the rented brain, the data you feed it, the agent sprawl problem, the incident you will eventually have, and the regulatory fog. With real-world examples and practical steps, it teaches you how to ask what your agents can actually do, how to limit what they touch, and how to build oversight before you need it. If you have ever clicked allow without reading, this funny, sharp, and genuinely useful book is the security review your inbox never had.