CISO warns against in-model AI controls

The author argues that AI safety controls embedded within a model's weights and context window are insufficient because they can be bypassed, making an independent enforcement layer necessary. The playbook outlines a hub-and-spoke gateway architecture featuring five mandatory control layers—identity, inspection, policy-as-code, response mediation, and immutable audit with a split-role kill switch—to govern AI interactions outside the model's environment. Open article →