Part III: Platform and service compromise

A model that passed evaluation can still be exposed through acquired software, shared infrastructure, or the identities and interfaces of the running service.

Downloaded models and packages become inputs to a running service. Their origin and integrity matter before deployment, but checks on those files do not establish that a host isolates tenants or that an interface limits access. Chapter 7 examines acquired components, Chapter 8 examines shared infrastructure, and Chapter 9 follows identities, network access, and resource use during operation. Internal and external model arrangements assign different controls and evidence to the organization and its providers.

Three large numbered panels show supply chain compromise, isolation on shared infrastructure, and runtime intrusion and resource abuse. The first panel ends at an admitted component, the second at isolation control and authorized decryption, and all three labeled outputs enter an organization workload in the third panel. Runtime arrows pass through an endpoint gate, scheduler, worker boundary, token, network filter, trace, containment, and service continuity.
Figure 1: Platform security starts by admitting a specific component, then establishing isolation and decryption conditions before runtime work begins. Component, isolation, and decryption decisions enter the workload through separate labeled ports, and runtime controls constrain requests, execution, network access, and evidence. A trusted platform therefore depends on linked admission, isolation, authorization, monitoring, containment, and continuity controls.

Chapters in this part

  • Supply chain compromise: A model file from the intended supplier can still execute unwanted code during loading or behave harmfully during use.
  • Isolation on shared infrastructure: Data processed by an AI service can be exposed or altered through the platform on which it runs.
  • Runtime intrusion and resource abuse: Exposed interfaces, stolen credentials, unrestricted network access, and excessive work can compromise an AI service or make it unavailable to legitimate users.