AB-100 governance controls
Security and governance questions usually come down to matching the control to the failure. This page covers what each of the five controls does, what it does not do, and where each belongs in the deployment lifecycle.
The Five Governance Controls
1. DLP (Data Loss Prevention) Policies
What it is: Power Platform policies that classify connectors into Business, Non-business, or Blocked groups. Connectors in different groups cannot be used together in the same app/flow/agent. Configured in the Power Platform Admin Center.
When to use it (AB-100 scenarios):
- Agent uses Dataverse + external connector (e.g., social media) and you need to prevent data leakage
- Regulated industry requires controlling which services can exchange data
- Agent could exfiltrate business data through an unapproved connector
- You need to prevent makers from combining sensitive connectors (e.g., Dataverse + Twitter)
What risk it mitigates: Data exfiltration, unauthorized data flows between services, accidental exposure of business data through unapproved connector combinations.
Scope: Environment-level (applied per environment or across multiple environments). Can also be tenant-level (applies to all environments). Environment-level DLP is the answer for scoped data protection within a regulated environment.
Common WRONG answers:
| Wrong Answer | Why It's Wrong |
|---|---|
| "Lower the temperature" | Temperature controls creativity/randomness of model output. It has ZERO effect on data flow or connector access. It is a model parameter, not a governance control. |
| "Configure tenant settings" | Tenant settings control broad admin capabilities (who can create environments, trial access). They don't control connector-to-connector data flow. DLP policies do. |
| "Use content filters" | Content filters (Azure AI Content Safety) block harmful content categories (hate, violence). They don't prevent data from flowing between connectors. |
| "Enable sensitivity labels" | Purview sensitivity labels classify and protect documents. They don't control which connectors can talk to each other. |
2. Managed Environments
What it is: An environment-level setting in Power Platform that enables additional admin controls: sharing limits, solution checker enforcement, maker welcome content, weekly digest emails, and the ability to require solutions for customizations. Toggled on per environment in the Admin Center.
When to use it (AB-100 scenarios):
- Production environment needs stricter controls than dev
- You need to enforce that all components are solution-aware (portable via ALM)
- Admins need to limit how many users a maker can share an app/agent with
- You want to require solution checker to pass before deployment
- Regulated environment needs visibility into maker activity (weekly digests)
What risk it mitigates: Uncontrolled maker activity, non-portable customizations, shadow IT, ungoverned agent proliferation, deployment of low-quality solutions.
Scope: Environment-level. Each environment is individually toggled to managed or default. This is the answer when you need environment-specific governance, not tenant-wide.
Common WRONG answers:
| Wrong Answer | Why It's Wrong |
|---|---|
| "Configure tenant settings" | Tenant settings are global. If a question says "in a specific environment" or "for the production environment," the answer is managed environments, not tenant settings. |
| "Use a default environment with restrictions" | Default environments cannot enforce solution checker, sharing limits, or maker onboarding. You must enable managed environment for these controls. |
| "Create a security group" | Security groups control WHO can access an environment, but don't enforce HOW they develop in it. Managed environments do. |
3. Microsoft Entra ID (Authentication & Authorization)
What it is: Microsoft's identity platform (formerly Azure AD). Provides authentication (who are you?), authorization (what can you access?), conditional access policies, MFA, service principals, and managed identities for agents.
When to use it (AB-100 scenarios):
- Agent needs to authenticate users before providing data
- Agent must respect user identity to enforce row-level security on Dataverse data
- Conditional access: block agent use from untrusted networks or non-compliant devices
- Service-to-service auth: agent calling an API needs a managed identity, not stored credentials
- SSO: user signs into Teams once and agent inherits their identity
What risk it mitigates: Unauthorized access, identity spoofing, credential theft, excessive permissions, unauthenticated agent use.
Scope: Tenant-level (Entra ID is configured at the tenant), but effects are agent-level (each agent configures its auth settings in Copilot Studio > Security > Authentication).
Common WRONG answers:
| Wrong Answer | Why It's Wrong |
|---|---|
| "Lower the temperature" | Temperature is a model parameter for output randomness. It does not authenticate users or control access. |
| "Add role mapping" (for prompt injection) | Role mapping controls authorization (what you can do). Prompt injection is an input attack that bypasses instructions. Role mapping doesn't block malicious prompts. Use prompt shields and input validation instead. |
| "Increase confidence threshold" | Confidence thresholds affect how the NLU matches intents. They have nothing to do with user identity or access control. |
4. Approval Gates
What it is: Workflow steps that require explicit human approval before an action proceeds. In Power Platform, these are Power Automate approval flows. In ALM, they are approval stages in deployment pipelines. In AI Foundry, they are model approval gates before production deployment.
When to use it (AB-100 scenarios):
- Agent performs high-risk actions (financial transactions, customer data changes)
- Deployment pipeline: solution must be approved before promoting to production
- Model registration: new model version needs human sign-off before deployment
- Agent autonomy boundary: autonomous agent needs a human-in-the-loop checkpoint
- Regulated workflow: compliance officer must approve before agent processes regulated data
What risk it mitigates: Unauthorized autonomous actions, unvalidated deployments, rogue model versions in production, compliance violations from unsupervised agent actions.
Scope: Agent-level (per agent or per flow) for runtime approvals. Pipeline-level for ALM deployment approvals. This is the answer when the question involves human oversight of agent actions or deployment gating.
Common WRONG answers:
| Wrong Answer | Why It's Wrong |
|---|---|
| "Enable monitoring/telemetry" | Monitoring observes what happened AFTER the fact. Approval gates PREVENT things from happening without sign-off. Monitoring is reactive; approval gates are proactive. |
| "Add a content filter" | Content filters block harmful content categories. They don't gate business decisions or deployment promotions. |
| "Use rollback" | Rollback undoes a deployment after it goes wrong. Approval gates prevent bad deployments from happening in the first place. Different phase of the lifecycle. |
5. RBAC (Role-Based Access Control)
What it is: Assigning permissions based on roles rather than individual users. In Copilot Studio: Viewer, Editor, Admin. In Dataverse: security roles with entity-level CRUD permissions. In Azure: built-in and custom roles (Reader, Contributor, Owner, etc.).
When to use it (AB-100 scenarios):
- Control who can edit vs. use an agent (Copilot Studio roles)
- Limit which Dataverse tables an agent can read/write based on the calling user's security role
- Separate maker permissions from admin permissions in an environment
- Control who can deploy solutions vs. who can develop them (separation of duties)
- Grounding data access: different users see different data through the same agent (security trimming)
What risk it mitigates: Excessive permissions, privilege escalation, unauthorized modifications to agents or data, insider threats, violation of least-privilege principle.
Scope: Multi-level. Tenant-level (Entra ID roles), environment-level (Dataverse security roles), agent-level (Copilot Studio sharing/roles), resource-level (Azure RBAC on AI Foundry resources).
Common WRONG answers:
| Wrong Answer | Why It's Wrong |
|---|---|
| "Role mapping" (for prompt injection) | RBAC/role mapping controls what a user CAN do. Prompt injection is an ATTACK on the agent's instructions. Even a user with correct permissions can attempt prompt injection. Use prompt shields, not RBAC. |
| "DLP policy" (for user access) | DLP controls connector data flows, not user-level permissions. If the question is about WHO can do WHAT, the answer is RBAC, not DLP. |
| "Lower the temperature" | Temperature does not restrict user permissions. Ever. |
Comparison Matrix
| Control | Scope | Configured In | Prevents | Does NOT Prevent |
|---|---|---|---|---|
| DLP | Environment / Tenant | PP Admin Center | Data flowing between wrong connectors | Unauthorized user access |
| Managed Env | Environment | PP Admin Center | Ungoverned maker activity, non-portable solutions | Prompt injection, data exfiltration via connectors |
| Entra ID | Tenant (agent-level config) | Entra Admin + Copilot Studio Security | Unauthorized users, unauthenticated access | Data flowing between connectors, bad model output |
| Approval Gates | Agent / Pipeline | Power Automate, ALM pipelines | Unreviewed actions, unvalidated deployments | Prompt injection, data leakage |
| RBAC | Multi-level | Entra, Dataverse, Copilot Studio, Azure | Excessive permissions, unauthorized edits | Model output quality, connector data flows |
Which Control for Which Incident? (Quick Reference)
| Incident Type | Correct Remediation | WHY This One |
|---|---|---|
| Unauthorized agent action (agent does something it shouldn't) | Restrict permissions (RBAC) | The agent had too many permissions. Tighten the Dataverse security role or Copilot Studio role. |
| Glossary/knowledge tampering (someone modified a managed solution's knowledge base) | Freeze the managed solution (lock it down, redeploy from source) | Managed solutions are immutable when imported. If tampered, the environment is compromised. Redeploy the clean managed solution. |
| Prompt injection attack (user tricks the agent into ignoring instructions) | Prompt Shields (Azure AI Content Safety) + input validation | Prompt injection is an input attack. The fix is input validation + prompt shields (Azure AI Content Safety), not temperature or RBAC. |
| Data residency violation (data processed in wrong region) | Force endpoint to compliant region | Configure the environment's data region and restrict Azure AI endpoints to the required geography. This is infrastructure config, not a model parameter. |
| Data exfiltration via connector (agent sends data to unapproved service) | DLP policy | Classify the connector as blocked or non-business. DLP prevents connector combinations. |
| Unauthorized user accessing agent | Entra ID authentication + conditional access | The user isn't supposed to be there. Fix identity, not data flow. |
| Unvalidated deployment to production | Approval gates in ALM pipeline | No human reviewed the deployment. Add an approval step before promotion. |
| Maker creating ungoverned agents in production | Managed environment | Enable managed environment to enforce solution checker, sharing limits, and maker controls. |
Things That Are NOT Governance Controls
Model Parameters (NEVER the answer for security questions)
| Distractor | What It Actually Does | Why It's WRONG for Security |
|---|---|---|
| Lower the temperature | Controls randomness/creativity of model output (0 = deterministic, 1 = creative) | Does not authenticate users, block prompt injection, prevent data leakage, or restrict permissions. It changes how "creative" the model sounds. That's it. |
| Increase/decrease confidence threshold | Controls how confident the NLU engine must be before matching an intent | Does not prevent unauthorized access, data exfiltration, or prompt injection. It affects intent matching accuracy. |
| Adjust top-p / top-k | Controls token sampling diversity during generation | Same as temperature. Model output tuning, not a security control. |
| Change the system prompt | Sets the agent's persona and instructions | Can be PART of prompt injection defense, but alone it is not a governance control. System prompts can be overridden by injection attacks. |
| Add few-shot examples | Improves model output quality by providing examples | Training technique, not governance. |
Operational Tools (NOT governance, even though they sound related)
| Distractor | What It Actually Does | When It IS Correct |
|---|---|---|
| Rollback | Reverts a deployment to a previous version | Correct for: "a bad deployment already happened, undo it." WRONG for: "prevent bad deployments" (use approval gates). WRONG for: "glossary tampering" (use managed solution freeze). |
| Enable telemetry/monitoring | Observes and logs agent behavior after the fact | Correct for: "detect anomalies, track KPIs, measure adoption." WRONG for: "prevent" anything. Monitoring is reactive. Governance is proactive. |
| Add logging | Records events for audit trail | Correct for: "compliance audit, post-incident investigation." WRONG for: "block unauthorized actions in real time." |
| Retrain the model | Updates model weights with new data | Correct for: "model accuracy is degrading, drift detected." WRONG for: any security incident. |
Governance in the deployment lifecycle
Correct Deployment Sequence
1. PLAN ─ Requirements, use cases, governance FRAMEWORK (policies defined here)
2. DESIGN ─ Agent architecture, security controls DESIGNED (DLP, RBAC, Entra, approval gates)
3. DEVELOP ─ Build in dev environment (managed environment enabled)
4. TEST ─ Functional, security, UAT testing
5. GOVERN ─ Approval gate, solution checker, compliance review ← BEFORE deployment
6. DEPLOY ─ Promote managed solution to production
7. MONITOR ─ Telemetry, logging, analytics ← AFTER deployment
8. OPERATE ─ Incident response, retraining, optimizationKey Ordering Rules
| Rule | Why |
|---|---|
| Governance BEFORE deploy | You must have approval gates, DLP policies, and RBAC configured BEFORE the solution reaches production. If governance is after deploy, you have an ungoverned production agent. |
| Telemetry AFTER deploy | You can't monitor what isn't deployed yet. Telemetry setup is part of the deployment config, but telemetry DATA collection starts after go-live. |
| Security controls DESIGNED, not bolted on | Security is designed in the DESIGN phase, implemented in DEV, validated in TEST, enforced via governance gates. It's not a post-deploy add-on. |
| Testing BEFORE governance gate | The solution must pass tests before the approval gate. Don't put approval before testing -- approvers need test results to make informed decisions. |
Regulated scope: environment DLP vs. tenant settings
Environment DLP vs. Tenant Settings Decision
| If the question says... | The answer is... | NOT... |
|---|---|---|
| "In a specific environment" / "for the production environment" | Environment-level DLP or Managed Environment | Tenant settings (too broad) |
| "Across all environments" / "organization-wide" | Tenant-level DLP or Tenant settings | Environment-level (too narrow) |
| "Control which connectors can be used together" | DLP policy | Tenant settings (wrong mechanism) |
| "Control who can create environments" | Tenant settings | DLP (wrong mechanism) |
| "Enforce solution checker before deploy" | Managed environment | Tenant settings |
| "Limit app sharing to N users" | Managed environment | DLP |
Embedded Chat vs. Generative Pages
| If the question says... | The answer is... | NOT... |
|---|---|---|
| "Dynamic, AI-generated interface based on data" | Generative pages (Copilot Studio) | Embedded chat |
| "Add the agent to an existing web page" | Embedded chat (iframe/widget) | Generative pages |
| "Agent creates rich, interactive responses with cards and data" | Generative pages | Embedded chat |
| "Simple Q&A on a website" | Embedded chat | Generative pages |
Memory Anchors
Use these to lock in the distinctions:
- DLP = connectors. If the question mentions connectors talking to each other, it's DLP.
- Managed Environments = maker controls. If the question mentions controlling how people BUILD, it's managed environments.
- Entra ID = identity. If the question mentions WHO someone is, it's Entra ID.
- Approval Gates = human checkpoint. If the question mentions PREVENTING something from proceeding without review, it's approval gates.
- RBAC = permissions. If the question mentions what someone CAN DO, it's RBAC.
- Temperature = NEVER. If temperature appears as an answer to a security question, it is wrong. Always. Without exception.