Pluto Copilot Studio reconciliation (2026-11-12)
Scope and baseline
Microsoft → Microsoft platform → Microsoft Copilot Studio, confirmed through Microsoft’s security and governance documentation. The existing Microsoft hub, Entra and Intune guides did not mention Copilot Studio. Baseline: zero product controls/packs; resulting guide: 14 controls, four with read-only automation, two implemented pack types. Existing Microsoft common controls are linked, not duplicated. No live tenant access or claim of ai-validated status.
Source-first reconciliation
The Plutonium safe-use index links three Microsoft articles. All three were fetched in this session. The former copilotsec.ai portal redirects to Plutonium; its current safe-use catalog points to these same three articles:
- Introducing CopilotSec: discovery/index and connector-review context, not a technical baseline on its own.
- Inside Copilot Studio: architecture/research context; model, delegation and artifact risks mapped to sections 2–4. Research-tenant defaults are not promoted to universal facts.
- Securing Copilot Studio: prescriptive candidate inventory below.
| Candidate theme | HTH disposition |
|---|---|
| Authentication/channels | New §1.1 and §2.1; primary-source enum correction in audit |
| HTTP passthrough | New §2.2; structured values still require validation |
| Delegated agents | New §2.4; completion options do not suppress child messages |
| MCP changes | New §2.3; tool allowlist and connector policy |
| Computer use | New §3.1, §4.2; separate execution and artifact boundaries |
| Sensitive routing | New §3.2, §3.4 |
| Tenant/environment policy | New §1.2, §2.1 |
| Opaque names | Not adopted as access control; avoid secrets in names, but obscurity cannot substitute for permissions |
| Retention and audits | New §4.1–4.3; metadata-only packs |
| Model/knowledge recommendations | New §3.3–3.4; current provider gates and grounding limitations |
Verified conflicts and bounded omissions
- Microsoft bot schema documents authenticationmode 0 Unspecified, 1 None, 2 Integrated, 3 Custom Entra ID, 4 Generic OAuth2. The audit preserves these values and flags unknown values; no heuristic is labeled proof of secure authentication.
- Microsoft data-policy documentation specifies real-time enforcement. Existing-agent testing remains required, but the guide does not claim enforcement is only point-in-time.
- Child-agent completion governs the parent’s completion message, not earlier child messages. A fixed completion is not a sanitization boundary.
- Runtime choices label both hosted browser and Cloud PC pools preview. Neither is described here as a universally production-ready substitute.
- Flow-session binary schema was verified for metadata reads. The research-specific
type='CuaScreenshot'filter was not confirmed in that current schema and is not shipped. No screenshot-delete script or universal default-role matrix is copied. - Broad tenant enumeration of Agent ID through Graph with no administrator privilege was not independently established. No enumeration-permission claim or opaque-name enforcement pack is shipped.
- Fixed provider/model versions, unconditional persisted chain-of-thought, specific default role depths, screenshot retention guarantees and universal prompt-injection bypass claims are research observations, not current guarantees established here. Validate effective permissions/artifacts with synthetic tenant tests.
Automation surface census
| Surface | Evidence and disposition | Read/write decision |
|---|---|---|
| Terraform | Microsoft provider resource confirms DLP resource; registry index returned a JavaScript shell and GitHub directory fetch was blocked. Complete resource/data-source enumeration remains incomplete, not absent. Releases flag tenant-setting deprecations. | Do not invent a comprehensive Terraform policy; scoped follow-up needed for schema/version-locked enforcement. |
| API | Current bot, botcomponent, flowsessionbinary schemas fetched; handles available. Paging contract verified. GraphQL control surface not verified. | Three GET-only audits; no direct mutation of agent internals. |
| CLI | Existing inventory had M365/Entra/Intune rows but no Copilot Studio. Official PAC command index confirms list/clone/pull/push/publish/quarantine. | Available pac copilot list --environment shipped; write commands deliberately not executed. |
| SDK | Dataverse ServiceClient is the current .NET client. | Same record plane as Web API; no redundant SDK pack or unverified setting setter. |
| Config | PAC workspace authoring exists; HTTP node and tool control prose verified. Full stable setting serialization for all controls not verified. | No fabricated YAML toggles; console implementation remains available. |
| SIEM | Purview support fetched; channel prerequisites apply. | No invented Sentinel/Splunk table or event schema; hub audit guidance linked. |
The two implemented pack types are intentionally read-only. This is not a claim that the other four surfaces are unavailable. Settings without verified setters carry explicit automation gaps in the guide.
Trust Center / baseline discovery boundary
Microsoft’s official security overview links the Microsoft Trust Center. Its product catalog includes Microsoft 365 and Power Platform. The Service Trust Portal returned a JavaScript shell; restricted document catalog completeness and government secure-configuration documents remain unverified. No gated report was downloaded and no assurance page was admitted as a configuration baseline. Public Microsoft Learn secure-project and governance guidance supplies the implementable controls above.
Bounded public searches for Copilot Studio CIS, STIG and SCuBA did not establish a dedicated prescription. This is not proof that none exists. Maester’s catalog surfaced Copilot Studio tests as a possible additional assessment lead; their individual implementation was not verified in this pass. Microsoft 365/Entra benchmark coverage is complementary.
Validation
Four shell scripts pass bash syntax checks. Eleven offline tests exercise pagination, exact authentication classifications, unknown enum handling, foreign-origin and looping links, error fail-closed, hash-only component output, invalid origins and metadata-only artifact queries. Fixture tests do not prove live permissions, API availability in every region/harness, or effective control enforcement. Raw component definitions are fetched only into a restricted temporary directory to calculate hashes, never printed; protect the machine and token accordingly.