Security PoC

Copilot Studio Agent Security

A clickable mockup on sample data showing how the PoC would control, inventory and check Copilot Studio agents. Nothing here is connected to a live Microsoft tenant.

Clickable mockup - sample data - based on public Microsoft documentation
1Control Lab
2Agent Inventory
3Risk Detection
4Our Approach
Runtime Security Architecture
Copilot Studio
AI Agent
→
Proposed
Action
→
External
Security Service
→
Policy
Evaluation
→
Decision
Allow / Block

Test an agent action

Select a scenario below and click Run to see a simulated trace of how the security service would evaluate and control the agent action.

Safe data read
Agent queries SharePoint list (read-only connector, authenticated user)
Allowed
Approved data write
Agent updates a CRM record through a governed Dataverse connector
Allowed
Blocked external email
Agent attempts to send email to an external address via Outlook connector
Blocked
Blocked privilege escalation
Agent tries to modify security group membership through Graph API
Blocked
Discovered agents
Sample data
Agent Type Environment Auth Risk Flags
Customer Support Bot Generative AI Production Entra ID SSO Low 3 connectors
Agent ID
bot_cs_001
Created
2026-09-15
Connectors
SharePoint (read), Dataverse (read), Knowledge base
Runtime hook
Supported (generative orchestration)
Last activity
2026-11-17 14:23 UTC
Owner
admin@contoso.onmicrosoft.com
Sales Pipeline Agent Autonomous Production Entra ID SSO Medium Data write5 connectors
Agent ID
bot_sp_002
Created
2026-10-02
Connectors
Dataverse (read/write), Outlook (send), SharePoint (read), Power Automate, SQL Server
Runtime hook
Supported (generative orchestration, event trigger)
Last activity
2026-11-18 09:41 UTC
Owner
sales.admin@contoso.onmicrosoft.com
HR Onboarding Assistant Classic Production No auth High No authPII access
Agent ID
bot_hr_003
Created
2026-06-20
Connectors
SharePoint (read/write), HTTP Request (custom)
Runtime hook
Not supported (classic agent)
Last activity
2026-08-14 11:02 UTC
Owner
hr.team@contoso.onmicrosoft.com
IT Helpdesk Copilot Generative AI Development Entra ID SSO Medium Maker creds
Agent ID
bot_it_004
Created
2026-10-28
Connectors
ServiceNow (read/write), Graph API (user read), Teams
Runtime hook
Supported (generative orchestration)
Last activity
2026-11-18 10:15 UTC
Owner
dev.admin@contoso.onmicrosoft.com
Finance Report Generator Generative AI Production Entra ID SSO High Over-privilegedData write
Agent ID
bot_fin_005
Created
2026-07-10
Connectors
SQL Server (read/write), SharePoint (full control), Power BI, HTTP Request (custom)
Runtime hook
Supported (generative orchestration)
Last activity
2026-11-17 16:50 UTC
Owner
finance.admin@contoso.onmicrosoft.com
Legacy FAQ Bot Classic Production Entra ID SSO Medium Dormant 90d+
Agent ID
bot_faq_006
Created
2025-12-01
Connectors
Knowledge base (read-only)
Runtime hook
Not supported (classic agent)
Last activity
2026-08-01 08:30 UTC
Owner
it.admin@contoso.onmicrosoft.com
Coverage note: Microsoft's external threat detection hook (preview) is only called for agents that use generative orchestration, including autonomous agents built that way. It is skipped for classic agents (Microsoft documentation), so classic agents get posture-based detection (inventory scan + configuration audit) instead of real-time blocking.
Configuration risk checks
Sample findings
Public access agent (no authentication)From your spec
Critical
An agent published to production without requiring user authentication. Any visitor with the link can interact with the agent and access connected data sources.
Detection method
Read each agent's authentication mode from the Dataverse agent (bot) table and the Power Platform inventory. Flag agents set to "No authentication" in production environments.
Remediation
Enable Entra ID SSO or manual authentication. Restrict to authenticated users through the agent's security settings in Copilot Studio.
Found in sample: HR Onboarding Assistant (bot_hr_003) - no authentication, PII-connected SharePoint site
Over-privileged AI userFrom your spec
Critical
An agent's service account or connection identity holds permissions beyond what the agent needs: Global Admin, write-all, full SharePoint control, or broad SQL access.
Detection method
Dataverse + Graph API: enumerate connector identities, map to Entra ID roles and resource permissions. Flag any agent with admin-level or broad write access.
Remediation
Apply least-privilege: scope connector permissions to the specific tables, lists or endpoints the agent needs. Remove admin roles from service accounts.
Found in sample: Finance Report Generator (bot_fin_005) - full SharePoint control + SQL read/write on production database
Autonomous data-write toolFrom your spec
High
An agent uses connectors that write, update or delete records in production systems (CRM, database, file storage) without a human approval step in the workflow.
Detection method
Scan agent connector permissions for write/update/delete operations. Cross-reference with workflow definitions to check for missing approval gates.
Remediation
Add approval flow before write actions, or use the external security service to require policy-based approval at runtime.
Found in sample: Sales Pipeline Agent (bot_sp_002) - Dataverse write + Outlook send with no approval gate
Dormant agent (90+ days inactive)Additional coverage
Medium
An agent in production has not been used in 90 or more days but retains active connectors and permissions. Dormant agents are often unpatched and unmonitored.
Detection method
Compare the agent's last recorded activity (audit log or Application Insights telemetry) with today. Dataverse transcripts are cleared after one month by default, so they cannot prove 90 days of inactivity on their own.
Remediation
Review and decommission or archive. If still needed, update configuration and re-test before reactivating.
Found in sample: Legacy FAQ Bot (bot_faq_006) - last activity 109 days ago, still published
Maker credentials in productionAdditional coverage
Medium
An agent's connector runs using the maker's personal credentials instead of a dedicated service account. If the maker leaves, the agent breaks. If the maker has admin rights, the agent inherits them.
Detection method
Check connection references for user-context sharing mode. Identify agents where the connection identity matches the creating user rather than a service principal.
Remediation
Switch to service principal authentication or a dedicated service account with scoped permissions.
Found in sample: IT Helpdesk Copilot (bot_it_004) - ServiceNow connector runs as the developer's personal account
Coverage by agent type: Agents on generative orchestration (including autonomous ones) get both posture scanning (inventory + config audit) and runtime interception (external security service). Classic agents get posture scanning only. The runtime hook must be enabled per environment by a Power Platform administrator.
Milestones
Milestone 1 - 1 week $750
Agent lab + API integration + risk detection
  • Build several agent types in Copilot Studio: classic topic-based, generative orchestration with tools, and autonomous with event triggers
  • One-command rebuild script using Power Platform CLI and solution export/import, after a one-time tenant setup (admin consent, enabling the hook) that is documented step by step
  • API integrations: agent inventory and configuration (Dataverse, Power Platform inventory), conversation transcripts (Dataverse), audit events (Microsoft Purview audit log)
  • 3 risk detection checks matching your spec + 2 additional, with capability matrix by agent type
  • Deliverable: working lab with data flowing and risks detected, full coverage documentation
Milestone 2 - 1 week $750
External action control service + handoff
  • External security service using Microsoft's custom-provider contract (POST /validate, POST /analyze-tool-execution)
  • Entra app registration with federated credential, policy evaluation engine, blockAction decisions inside Microsoft's one-second limit, with error behavior set to Block the query so a timeout fails closed
  • Coverage: agents on generative orchestration intercepted at runtime; classic agents covered by posture scanning
  • Proof of a block at the destination (the write never lands), always paired with an allowed control test
  • Complete handoff documentation: architecture, API reference, risk rule specs, deployment runbook
  • Independent rebuild verification: a second engineer runs the rebuild from documentation alone
$1,500 total
Fixed price - 2 weeks - remote - start immediately
No data loss
No downtime
Nothing breaks
Pay on acceptance
Your questions answered
Tell us about a Copilot Studio project you have worked on
Our Microsoft 365 work so far has been on the application side: a Purview eDiscovery evidence-check app that runs on synthetic data, and an AI admin layer built on Microsoft Graph and Power Automate. Our security team has built enterprise governance frameworks and authored 30+ CVE deep-dive security research pieces. For this post we worked through the Copilot Studio threat detection contract; this page is a clickable mockup of that work on sample data, with no live tenant connection.
Which part of the job are you strongest at?
The API integration layer and the external action control service. Our backend architect has 16+ years in .NET and cloud architecture, with token and role-based access systems that passed independent security audits. Our security consultant (CISA, CISM, CRISC, ISO 27001 Lead Auditor) designs the detection rules. Together, they cover both the engineering plumbing and the security logic your product needs.
Do you have access to an AI-enabled Microsoft environment?
We can author and test in a Copilot Studio trial right away. A trial cannot publish agents, so the live hook test needs a licensed or pay-as-you-go environment: your PoC tenant, or one we set up with the Microsoft costs billed separately. In your tenant we would need a licensed (non-guest) account, admin consent for the Entra app registration, Power Platform environment access, and the external threat detection setting turned on. We start the day access is confirmed.
Our team
Security and GRC
About 10 years
CISA, CISM, CRISC, ISO 27001 Lead Auditor, C|CISO, PMP, Security+, CySA+. Risk assessment, secure code review, compliance programmes.
Backend and Cloud Architecture
16+ years
.NET Core, AWS (Certified Solutions Architect, Developer, Data Engineer). Token and role-based access systems that passed external security audits.
DevOps
AWS Professional Certifications
Solutions Architect Professional, DevOps Engineer Professional. Cloud-native deployment, monitoring, CI/CD.
QA Automation
13 years
Playwright, Selenium, Cypress, JMeter, API testing. Verifies every detection rule and integration path.
Questions for you
1. Which Microsoft 365 tenant edition and Copilot Studio license is the PoC running on, or should we set up a licensed environment on our side?
2. Are sample agents already built, or do we build the full agent set from scratch? Any existing assets to preserve?
3. For the action control service: which specific agent action should the first test prevent (a data write, an email send, a connector call)?
4. Does "rebuildable with one command" mean a rebuild into a prepared environment, or must a fresh environment also be fully automated?
5. Should documentation include API reference specs for your product team, operational runbooks, or both?
About this page

This is a clickable mockup on sample data, built from Microsoft's public documentation for the Copilot Studio external threat detection contract. The trace on the first tab is simulated and nothing here calls a live service. The paid engagement builds the real thing: the allow/block service, the API integrations, your test tenant and a documented handoff your team can run on its own.