Shadow AI: The Operator's Playbook for Finding the AI Your Team Already Uses and Governing Where Your Data Goes, Without a Security Team (Paperback) (ISBN-13: 9798188345693)

On March 11, 2023, Samsung's semiconductor division told its engineers they could start using ChatGPT. Within about twenty days, three confidential leaks followed: chip-measurement source code, defect-detection code, and the contents of an internal meeting, each sent to a server Samsung had no contract with and no way to delete.

None of it was sabotage. Three engineers used a tool their employer had just approved, to finish real work a little faster. That is the distinction this book runs on. The data was not at risk of leaking. It had already left, and nothing on any screen said so, because a copy leaves no gap the way a theft does. Samsung's first move was a 1,024-byte cap on every prompt, then a full ban seven weeks after saying yes. Neither answered what had gone out, or how anyone would know if it happened again.

Shadow AI is the loss-prevention playbook for the person who actually owns this exposure: a team lead or operations manager whose people are already pasting real work into tools nobody approved, on accounts nobody can see, with no security team to hand it to. It borrows an idea from retail loss prevention, shrink, merchandise on the books but gone from the shelf, and applies it to data that never technically goes missing. You cannot bolt every door. Banning the tool trades a risk you can see for the same risk, invisible. The book trades one unanswerable question, "Is this tool safe?", for four you can actually answer: where the data goes, who keeps it, whether you can see who is using what, and whether you can stop it.

It is a working kit, not a survey. Across 16 chapters you build the four artifacts that do the work, plus the reference you want open during an incident:

• The discovery survey. Finds the unapproved tools and personal accounts your admin console cannot see, the invisible majority of your team's real AI use.
• The data-sensitivity decision tree. Tells you in seconds what is safe to send, what to strip first, and what must never leave, before anyone pastes.
• The approved-tools list. One sanctioned tool per job, its data posture written next to it: a living go or no-go record, not a memo nobody reopens.
• The incident and leak response one-pager. Runs the first hour and the first 72 hours of a leak: who to tell, by when, what to try to get back.
• The vendor data-handling reference. What OpenAI, Anthropic, Microsoft, and Google say in their own current documentation about training defaults and retention windows, dated and re-verifiable.

None of it requires a security team. It requires the four questions, honesty about what a ban actually does, and sorting what is sensitive before anyone hits send. The Samsung account that opens the book runs back through the finished kit in the last chapter, closing on an ordinary Tuesday instead of a ban.

This is a volume in The Operator's AI Library: for team leads and operations managers whose people already use AI on accounts they cannot see. Readers of the author's Verifier's Library will recognize the discipline, especially Read the Halt and Sounds Right; those books ask whether an AI's answer can be trusted, this one asks where the data went before anyone saw the answer.