Microsoft adds task-scoped permissions and containment for AI agents

November 9, 2026

Microsoft’s November 2026 Security updates add new controls for organizations deploying AI agents, including task-scoped permissions, generally available Execution Containers, and discovery of local agents and their configured model context protocols (MCPs).

Permissions that can change with the task

Microsoft says Dynamic Permissions gives an agent authority limited to its current task and only for as long as that task requires. Dynamic Agent Risk evaluates whether the agent’s permissions still match its behaviour and context. Dynamic Conditional Access can then narrow access, restrict resources, or require human approval when the risk or impact increases.

The important shift is from static entitlement to conditional authority. A service account that can read a database or send a message may be acceptable for a narrowly defined workflow, but the same access becomes harder to justify when an agent changes tasks or encounters unfamiliar data. Microsoft’s approach is intended to make those boundaries adjustable while the agent is operating.

Execution Containers are now generally available

Microsoft Execution Containers provide an operating-system-enforced containment layer for AI agents on Windows. Microsoft describes controls across files, networks, processes, and user interactions, with additional isolation options for higher-risk scenarios.

That matters because policy cannot depend entirely on an agent following instructions. A container can limit what the software is technically able to reach, reducing the consequences of a faulty prompt, compromised tool, or unexpected action. Microsoft says organizations can tune and validate policies before enforcing them broadly. Planned integration with Intune, Microsoft Entra, and Microsoft Agent 365 is intended to extend governance and attribution across enterprise environments.

More visibility into local agents and MCPs

Microsoft Defender’s local-agent discovery is now generally available for supported agents on Windows and macOS. Microsoft says Defender can identify more than 50 supported local AI agents and combine information about the agent, device, account, and configured MCPs.

MCP is a protocol that allows an AI application to connect to external tools and data sources. Inventory alone does not make those connections safe, but it gives security teams a starting point for finding unmanaged agents and understanding which systems they can reach.

What security teams should take from the update

  • Inventory agents, owners, devices, tools, and data access before approving new workflows.
  • Prefer least privilege that is scoped to a task instead of broad, permanent permissions.
  • Use containment for local agents that can read files, invoke tools, or interact with users.
  • Require human approval for actions with financial, identity, production, or external-communication impact.

These are Microsoft announcements, not a claim that every capability is available in every tenant or licensing plan. Organizations should confirm regional availability, prerequisites, and integration status before building a control framework around them.

Microsoft’s announcement also connects the changes to its Secure Now exposure-management work, which maps recent threat information to foundational security initiatives. For businesses already evaluating AI-agent governance, that broader context is more useful than treating agent security as a separate product feature.

Source: Microsoft Security: What’s new in November 2026

Related: Cloudflare adds traffic analytics and access rules for individual AI agents.

Feature image: AI-generated conceptual illustration for Guava Insights.

Topics:
Read next
November 21, 2026

How to Require MFA for Administrator Roles with Microsoft Entra Conditional Access

Protect privileged Microsoft Entra accounts by creating a report-only Conditional Access policy that requires MFA for built-in administrator roles, then safely moving it to On.
November 20, 2026

U.S. Expands AI-Chip Controls to Remote Cloud Access as China Pushes Domestic Alternatives

A new U.S. interim rule extends advanced AI-computing controls to remote access through overseas data centres, raising compliance and infrastructure questions for cloud providers and Chinese technology companies.
November 20, 2026

OpenAI Releases GPT-6.2 Astra and Sol as AI Competition Shifts to Workflow Economics

OpenAI has released GPT-6.2 Astra and Sol with lower API pricing and stronger professional-work claims. Early independent reactions are mixed, making task-level testing more important than benchmark headlines.