What is an AI containment breach?

Containment is the set of controls that keep an agent's actions inside the scope its operator intended: the network boundary, the credentials it holds, the human approval steps, the monitoring while it runs, the log it cannot edit, and a way to stop it. A breach is an action that got past one of them. The Escape Record scores every documented case on eight such controls, so a breach is recorded not only as what happened but as which control was absent when it did.

The word is useful because production systems breach containment too. A deployed assistant that emails a customer internal notes it was never meant to send has exceeded its boundary as surely as an evaluation model that reached another company's servers. The Escape Record marks production cases as overreach and keeps the inclusion question open where the case is borderline, because a record that only counts laboratory incidents understates the shape of the problem.

What a breach is not: a hallucination, an ordinary bug, or a model refusing to do something. Those are failures of output. A breach is a failure of boundary.