Apply data policies to every environment, keep audit logs in Purview, check where data rests and which models process it, and test every agent against prompt injection before it reads real data.
When you need this
- An agent will read customer or employee records through connectors.
- Security asks how agents are audited and where their data lives.
- An agent will accept input from email, a website or another channel outside the tenant.
Evaluation criteria
Data policies
Which connectors can each agent use?
Look forPower Platform data policies apply to all tenants with no agent exemption. Block connectors agents do not need.
Audit
Is agent activity recorded?
Look forCopilot Studio writes audit logs to Microsoft Purview; review them with your security team.
Data location and models
Where does agent data rest, and which models process it?
Look forData stays in the tenant's geography, or the United States when the tenant's location is not listed. Models marked cross-geo can process data outside the region.
Prompt injection
Can an outsider steer the agent?
Look forZenity Labs showed an email-triggered agent connected to Salesforce exfiltrating data, Tenable got an agent to leak card data, and Datadog Security Labs showed agent pages used to phish OAuth tokens. Test each channel that can reach your agent.
Encryption and compliance
Do you need customer-managed keys or a compliance program?
Look forCopilot Studio supports customer-managed keys for environment data, and Microsoft lists its ISO, SOC and HIPAA coverage.
Questions to ask vendors
- Which agents accept input from outside the tenant?
- Which connectors does each agent hold, and does it need them all?
- Which models are allowed in each environment?
- Who reviews Copilot Studio audit logs?
Red flags
- An email-triggered agent with write access to a CRM.
- Cross-geo models allowed for data that must stay in region.
- Agents shared to the whole tenant by default.
- No prompt injection test before launch.
Products in this piece
Sources
- Official documentation
- 7
- Public source
- 3
- Microsoft Learn: Configure data policies for agents
- Microsoft Learn: View Copilot Studio audit logs in Purview
- Microsoft Learn: Data locations in Copilot Studio
- Microsoft Learn: Select a primary AI model for your agent
- Microsoft Learn: Customer-managed encryption keys for Copilot Studio
- Microsoft Learn: Review ISO, SOC, and HIPAA compliance
- Microsoft Learn: Security and governance in Copilot Studio
- Zenity Labs: A Copilot Studio story 2, when AIjacking leads to full data exfiltration (Jul 7, 2025)
- Tenable: Microsoft Copilot Studio security risk, how simple prompt injection leaked sensitive data (Dec 11, 2025)
- Datadog Security Labs: CoPhish, using Microsoft Copilot Studio as a wrapper (Oct 20, 2025)