Defra Logo Defra Logo

Defra AI in the Tool Guidance

Guidance on the use of artificial intelligence (AI) tools within the Department for Environment, Food & Rural Affairs (Defra), covering best practices, security considerations, and implementation guidelines.

Azure AI Foundry - Detailed Guide

Approval status: Approved - you can use this tool within Defra when you follow the tool guidance advice.

(Generated by AI, ChatGPT Deep Research, on April 25th June 2025)

1. Tool Overview

Azure AI Foundry is Microsoft’s managed platform for designing, building and running generative AI applications at scale.
It brings together model catalogues, multi-agent toolchains, managed compute and governance capabilities in a single portal, SDK and REST API.1

Key use cases include:

  • Evaluating and deploying foundation, open-source and domain models
  • Fine-tuning and safeguarding models with responsible AI tooling
  • Orchestrating multi-step “agent” workflows
  • Governing AI projects through hubs, projects and central policy controls

Website: https://learn.microsoft.com/en-gb/azure/ai-foundry/

2. Privacy Settings

  • Private networking – you can create hubs and projects inside a managed virtual network and access them exclusively through Azure Private Link, blocking public internet traffic.2
  • Customer-managed keys (CMK) – administrators can encrypt metadata and artefacts with keys held in Azure Key Vault.2
  • Conditional Access and RBAC – Microsoft Entra Conditional Access policies and Azure RBAC (role-based access control) roles restrict who can view or edit resources.3
  • Optional diagnostic data – administrators choose whether to stream resource logs to Log Analytics; no telemetry is enforced.4

3. Terms of Use and Privacy Policy

| Document | Purpose | Link | |———-|———|——| | Microsoft Product Terms | Governs use of all Azure services, including AI Foundry | https://www.microsoft.com/licensing/terms | | Microsoft Products & Services Data Protection Addendum | Contractual GDPR compliance and data-processing terms | https://aka.ms/dpa | | Supplemental Terms for Azure Previews | Applies to preview features in AI Foundry | https://azure.microsoft.com/support/legal/preview-supplemental-terms | | Microsoft Privacy Statement | Corporate privacy notice | https://privacy.microsoft.com/privacystatement |

4. Data Management

4.1 Server Location and Data Residency

  • Hubs, projects and agent endpoints are regional Azure resources. Runtime data stays in the same Azure region you select (for example UK South, UK West).5
  • Azure’s Europe geography meets UK data-residency requirements and is assessed against the NCSC 14 Cloud Security Principles.6

4.2 Data in Transit

  • All traffic uses TLS 1.2 or higher by default; you can block weaker cipher suites through policy.7
  • Private Link keeps traffic on the Microsoft backbone, removing exposure to the public internet.2

4.3 Data at Rest

  • Service-side encryption with AES-256 is always on.8
  • CMK (customer-managed keys) is supported for hubs, projects, agent threads and file storage.29

4.4 Data Retention

| Data type | Default retention | Administrator controls | |———–|——————|————————| | Agent Service threads and files | Persistent until caller deletes | Delete API | | Application Insights logs | 90 days | 30–730 days configurable10 | | Resource logs (Activity or Diagnostic) | 90 days | Azure Monitor retention policies | | Model inputs and outputs | Not used for training; stored only in customer subscription unless logged through above features1 |

5. Audit Logging

  • Resource logs for hubs and projects capture create and delete operations and can be exported to SIEM (security information and event management) tooling.4
  • Azure Monitor and Log Analytics provide metric and query-based alerting.11
  • Built-in Azure Policy definitions enforce that diagnostic settings are enabled and check compute resources for compliance.4

6. Access Controls

  • Azure RBAC roles such as Hub Admin, Project Contributor and custom roles govern least-privilege access.3
  • Managed identities (system or user-assigned) avoid embedded secrets when Foundry interacts with other Azure services.2
  • You can disable public network access at hub level; enforcement is possible through Deny policy.4

7. Compliance and Regulatory Requirements

Azure AI Foundry inherits Microsoft Azure’s broad certification portfolio, including:

Standard or Scheme Status Notes
ISO/IEC 27001:2022 Certified Certificate available through Microsoft Service Trust Portal12
SOC 2 Type 2 Attested Security, availability, processing integrity, confidentiality13
GDPR Compliant Contractual commitments in DPA
HIPAA and HITRUST Eligible workloads BAA available on request14
UK G-Cloud (NCSC 14 CSP) Assessed White-paper mapping Azure controls to 14 principles6

Note: Responsibility for end-to-end compliance is shared. You must configure policies, retention and key-management to match your risk profile.

8. References

  1. Data, privacy and security for use of models through the model catalogue, Microsoft Learn, 19 May 2025. ↩ ↩2

  2. Policies for Azure AI Foundry hubs and projects, Microsoft Learn, 11 Jun 2025. ↩ ↩2 ↩3 ↩4 ↩5

  3. Role-based access control in Azure AI Foundry portal, Microsoft Learn, 4 Jun 2025. ↩ ↩2

  4. ibid., built-in policy definitions for logging and network controls. ↩ ↩2 ↩3 ↩4

  5. Azure AI Foundry Agent Service FAQ, Microsoft Learn, 2025. ↩

  6. UK G-Cloud – Azure Compliance, Microsoft Learn. ↩ ↩2

  7. Enforce a minimum required version of TLS for Storage, Microsoft Learn, 22 Mar 2024. ↩

  8. Azure Storage encryption for data at rest, Microsoft Learn, 13 Feb 2023. ↩

  9. Customer-managed keys, Microsoft Learn (linked from policy doc). ↩

  10. Customer-enabled disaster recovery, Microsoft Learn, Apr 2025. ↩

  11. Azure AI Foundry architecture, Microsoft Learn, 28 Apr 2025. ↩

  12. ISO/IEC 27001 – Azure Compliance, Microsoft Learn. ↩

  13. System and Organisation Controls (SOC) 2 Type 2, Microsoft Learn. ↩

  14. HIPAA – Azure Compliance, Microsoft Learn. ↩

Support links

All content is available under the Open Government Licence v3.0, except where otherwise stated