Secure AI Development
This section is about the machine a developer writes code on, and the assistant that now runs there. It covers what an organisation can enforce on an endpoint it manages, what it can only observe once an assistant was installed by an individual, and which settings decide what reaches the model and what comes back with the authority to act.
The controls sit around the model, not inside it
What this section covers
The threat model sets out why an editor with an assistant is a boundary at all, and which control sits between untrusted repository text and an executed command. The four pages after it are the stack: the editor's policy surface, how that policy reaches a device, the repository platform that governs AI features and the merge gate, and the assistants no central policy reaches. The last page is the operating model.
The control stack
Each layer can enforce something the layers above it cannot, and each has a documented limit.
| Layer | What can be enforced there | Page |
|---|---|---|
| Device policy | Which editor build is present, whether extensions may install, and the telemetry level. | 03 — GPO and Intune |
| Editor policy | Updates, the marketplace, AI feature switches, permitted MCP servers, and the default approval mode. | 02 — VS Code as an endpoint |
| Assistant configuration | Per-assistant config files, permission rules that survive a saved approval, and the model provider it talks to. | 05 — OpenCode and IDEs |
| Repository platform | Which AI features are enabled, secret scanning and push protection, and the ruleset that decides what may merge. | 04 — GitHub Enterprise |
| Pipeline | Which workflow files may run, what token a job receives, and what a required check verifies. | 04 — GitHub Enterprise |
| Identity, network and data | Who may sign in from what device, which model endpoints are reachable, and what data controls sit on the endpoint. | 05 — OpenCode and IDEs |
Read it as an order of preference and a statement of dependency, not a checklist. Device policy cannot repair a merge gate that accepts anything, and a ruleset cannot stop a secret being pasted into a chat window. Where the first three rows do not exist, because the assistant is an individual's choice of tool, the last row is what remains.
What this section does not cover
- Model training security, model internals, and the evaluation of a model's own behaviour.
- Prompt engineering, and prompt-level defences presented as a substitute for an approval gate.
- Attacks on AI-enabled systems and the payloads for them. Those are in Offensive AI, and this section links to them instead of repeating them.
- Hardening of hosted AI services and their APIs, which belong to Cloud Pentesting and API Security.
- Legal advice on vendor terms. These pages say which page of a provider's documentation answers a question, not what it will say when you read it.
- Any claim that a control here satisfies a named framework requirement. The mapping is to an outcome you can demonstrate, not an attestation.
Related Sections
Secure Architecture
Design-time controls, threat modelling and trust boundaries for what is being built.
CI/CD & Supply Chain
The pipeline side of the commit boundary this section meets at the developer machine.
Source Code Review
Review method for code a human did not type, including AI-suggested changes.
Offensive AI
The attack taxonomy against assistants and agents that this section refuses to re-teach.
Threat & Risk Assessment
Framing residual risk where a control cannot exist.