Developers and system operators analyse threats to their systems and manage the resulting security risks. Threat models are reviewed and updated on a regular basis and cover attacks peculiar to AI (poisoning of training data, inversion of the model, inference of membership); the process is re-run whenever a setting or a configuration option is newly introduced or changed at any point in the lifecycle. Developers manage the extra risk carried by models with functionality beyond what the system uses (a multi-modal model used for a single modality, for instance), and system operators apply controls to the risks found, weighing implementation cost against the organisation's risk tolerance.
The graph holds this control, the 0 it maps to, and the evidence behind each claim, over MCP and REST.