Organisational policy that defines the lifecycle of the mandate of each AI agent in operational use — registration, ownership, autonomy levels A0-A4, tool scope, periodic review and revocation, in alignment with REQ-AGN-001..004 of Ch. 02 and ARC-015 of Ch. 04, proportionate to the criticality level (L1, L2, L3).
Organisational policy that defines the treatment of AI models, datasets, MCP servers/tools, embedded prompts and AI providers as an auditable supply chain — AI BOM generation per build in a standardised format (CycloneDX 1.6 ml-bom), version pinning, a list of approved providers and response to upstream incidents, complementing Policy 10 (dependencies) and Policy 11 (SBOM), proportionate to the criticality level (L1, L2, L3).
Cross-cutting map of the secure development process for systems with AI agents — from mandate to kill-switch, with stops in each SbD-ToE chapter and a cross-reference to the AI Act, NIS2, DORA and the CRA
Operational patterns for serving self-hosted AI models (vLLM, Ollama, TGI, llama.cpp, NVIDIA Triton) — workload isolation, weight management, protection of inference, AI-specific container hardening.
Good practices and controls for the secure use of generative AI tools in code writing and automated review