Module 3 of 5
Securing each stage
An AI application passes through seven stages, from design to operation. Each has controls that are cheap there and expensive anywhere else.
An AI application passes through seven stages: design, data, model, build, deploy, runtime, and operate. Each attack from the previous module has one stage where it is cheapest to stop, and each stage has controls that are cheap there and expensive anywhere else. This module walks the lifecycle in order, one lesson per stage, after a first lesson that maps the whole of it.
By the end you should be able to take any threat in a threat model, name the stage where it is cheapest to stop, place a deterministic control there, and say who owns it. You should also be able to tell which of your controls only make attacks rarer and which actually make them impossible.
Agents add enough new surface (identity, sandboxes, tool protocols) that they get their own module after this one: Securing agents.
Lessons
- 3.aThe lifecycle mapWhich control belongs at which stage, and who owns it?15 min
- 3.bData and trainingHow do I keep training, fine-tuning, and retrieval data trustworthy and private?14 min
- 3.cModels and supply chainHow do I know a model file is what it claims and safe to load?15 min
- 3.dBuild and testWhat goes in CI for an AI application, and how do I red-team it?16 min
- 3.eDeployment and infrastructureHow do I isolate, harden, and protect the serving environment?18 min
- 3.fRuntime guardrailsWhat do input and output checks achieve, and where do they stop?15 min
- 3.gMonitoring and responseHow do I detect an attack on an AI system and respond to it?15 min