[ ABORT TO HUD ]
SEQ. 1
SEQ. 2
SEQ. 3
SEQ. 4
Enterprise Governance & Security
Securing Agent Deployments
As agents gain autonomy, security transitions from static checks to runtime evaluation. Microsoft Foundry provides comprehensive governance mechanisms.
Microsoft Purview for AI
Microsoft Purview now integrates directly with Foundry, offering Runtime DLP (Data Loss Prevention) for agent prompts. It actively scans agent inputs and outputs to prevent PII leakage and enforce compliance policies in real-time.
AgentGuard
AgentGuard serves as the central security gateway for all outbound agent actions, ensuring that tool calls and API requests adhere to the principle of least privilege.
September 2026 Agent Security Policies
- Mandatory Endpoint Encryption: All agent tool calls and data grounding requests require TLS 1.3 encryption and must traverse Azure Private Link, eliminating public internet traversal.
- Granular Enterprise Bing Grounding DLP: Dynamic domain allowlisting prevents agents from leaking intranet context during web queries while restricting web citations to pre-approved corporate domains.
- Microsoft Purview AI Hub Integration: Real-time auditing tracks agent execution traces, sensitive data egress, and user prompt safety compliance across all Foundry projects.
Microsoft Foundry SOE (Standard Operating Environment)
| Component | Enterprise Baseline | Hardening Standard |
|---|---|---|
| SDK Baseline | azure-ai-projects >= 1.0.0b6 | Pinned dependency with weekly vulnerability scan; deprecate legacy v1 clients |
| Authentication | DefaultAzureCredential | Managed Identity with User-Assigned RBAC; zero stored connection strings or API keys |
| Network Topology | Azure Virtual Network + Private Endpoints | Deny public network access to Hub, Projects, and Azure AI Search resources |
| Model Governance | Catalog Version Pinning | Pin explicit semantic model version numbers; prohibit dynamic latest floating tags |
Microsoft Foundry AOE (AI Operations Engineering) Blueprint
- Distributed Tracing in Application Insights: Route all agent execution runs through Azure Monitor with custom OpenTelemetry spans tracking token latency, grounding recall, and tool evaluation status.
- Automated CI/CD Safety Gates: Run automated ASSERT test suites in GitHub Actions / Azure DevOps; block releases failing prompt injection or domain-specific safety checks.
- Prompt Shield Telemetry & Circuit Breakers: Trigger alert thresholds when
ContentSafetyFilterTriggeredrate exceeds 1% in a rolling 10-minute window; automatically fall back to restricted agent modes. - PTU Capacity & Burst Optimization: Monitor Provisioned Throughput Unit utilization and configure automated overflow routing to Global Standard endpoints during peak spikes.
FOUNDRY VERIFICATION
QUERY 1 // 2
What does the September 2026 Agent Security Policy mandate for agent tool invocations?
Public API keys
TLS 1.3 encryption and Azure Private Link network isolation
Manual human approval for every token
Disabling search tools entirely