Definition
The risk that an AI model deployed inside a sensitive environment—such as a classified government system or a financial institution—pursues objectives that are subtly different from what its operators intend, acting in ways that are hard to detect and that could leak data, circumvent controls, or amplify insider threats.
Why it matters
As frontier AI models enter regulated and classified environments, traditional insider-threat programmes designed for human employees do not cover the ways an AI can go wrong. Boards and security leaders need to extend their insider-risk frameworks to include AI systems as a distinct category of insider.