The release assistant requested a credential export. Did it leave?
FOR CISOs & SECURITY LEADERS
Know which agents can act—and what keeps them within policy.
Keep sensitive actions inside their allowed scope. Give security teams a clear account of what was requested, blocked and recorded.
Microsoft TeamsSecurity operations
Team conversationThe configured export path denied the request.
Export blockedBlocked
- Scope
- Deployment status read
- Credential export
- Denied before configured operation
- Effect
- No export receipt
Keep its scope unchanged and investigate the request source.
The exact request, actor, policy revision and denial are attached for investigation.
Request rejected. No action was dispatched. A new request needs its own review.
Ready
A credential request stops at the boundary.
An agent may inspect an approved release signal. A credential export must be checked before any tool sees it.
Read the decision path
- Security lead
- Asks about an export.
- Release agent
- Requests credential export.
- DriftGate
- Export policy checked.
- Denied before tool
- Credential export denied.
- Export tool
- Not reached.
- Evidence
- Denial and result recorded.
THE OPERATING MODEL
Make the work repeatable. Keep the authority explicit.
Start with one consequential workflow: a customer refund, production change, financial export or employee offboarding. Identify the trusted actor and protected resource. Demonstrate the allowed path, then test an unauthorized request, revoked grant and uncertain result.
Evaluate what is connected, who owns it, where permission is enforced, what happens when authority changes and which records support an investigation. A registered inventory is not automatic discovery of every unconnected shadow-AI system.
IN PRACTICE · EXAMPLE WORKFLOW
Read-only should be a permission. Not a request.
A coding assistant investigates a production problem but attempts a destructive database operation during diagnosis.
Separate development and production credentials. Grant read-only diagnostics, expose permitted typed operations and enforce policy at the executor. Restrict shell access so it cannot bypass the protected path.
See the workflowTHE INTENDED RESULT
Approved log reads succeed. Destructive production mutation is denied before dispatch instead of merely discouraged in a prompt.
SEE IT IN PRACTICE
Put the platform to work.
SOFTWARE
Read-only should be a permission. Not a request.
A coding assistant investigates a production problem but attempts a destructive database operation during diagnosis.
MCP & tool governance
FINANCIAL OPERATIONS
A convincing request is still just a request.
A caller asks a voice or chat assistant for a financial-data export and claims to be the CFO. The request sounds urgent and plausible.
Agent IAM
SECURITY OPERATIONS
Contain the incident without shutting down the company.
An automated responder proposes broad account suspension or a network change after detecting suspicious activity.
Agent lifecycle
SAME FOUNDATION
Explore the connected controls.
9 CAPABILITIES
Environment & Tooling Management
Bring tools, skills, instructions and configuration into one approved environment. Keep the right setup with the right project—not scattered across every client and computer.
Explore MCP & tool governance10 CAPABILITIES
Identity, Access & Policy
Give agents and workloads the access their tasks need—not unrestricted access to your business. Connect every governed request to an accountable identity, resource scope and current policy.
Explore Agent IAM10 CAPABILITIES
Runtime Governance & Execution Control
Put authorization, approvals and limits where your agents act. Follow each consequential operation from intent to an attributable outcome—even when the work waits or a response is lost.
Explore Runtime controls10 CAPABILITIES
Observability, Audit & Assurance
Connect identities, configurations, decisions and outcomes into an investigation path. Know what changed, why an action was allowed and which evidence supports the result.
Explore Audit & investigations10 CAPABILITIES
Lifecycle & Change Management
Give agents, connections, policies and knowledge a deliberate lifecycle. Review changes, release with intent, understand affected consumers and retire access without losing the history.
Explore Agent lifecycle10 CAPABILITIES
Data & Context Governance
Give agents the information their tasks need—not every document your organization owns. Keep source ownership, context boundaries and outbound data rules connected to the work.
Explore Data governanceStart with a sensitive operation, then test the boundary.
Map the workload identity, owner, connected resources and permitted operations. Ask whether the integration can only observe, can warn, can hold for approval or can enforce before the effect. Preserve the policy revision and decision alongside the resulting receipt.
Test unauthorized input, revoked authority and an uncertain provider outcome in a qualified environment. A registered inventory improves connected-system visibility; it does not establish discovery of every unconnected shadow agent or prevention on every external path.
BUILD WITH INTENTION